Examenos

Text Decision Index 0.3.2 · Full score

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Text Decision Index 0.3.2 · Public score

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Text Decision Index 0.3.2 · Private same-skill score

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Text Decision Index 0.3.2 · Private new-domain score

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

HF text 0.3.2 · Raw balanced index

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

Source-published alternative aggregate, separate from Full and Public scores; no Examenos recomputation.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published aggregate observations; result reuse terms not separately specified

HF text 0.3.2 · Public breadth index

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

Source-published alternative aggregate, separate from Full and Public scores; no Examenos recomputation.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published aggregate observations; result reuse terms not separately specified

Calibration on the source-defined text sample; confidence alone is not a quality ranking.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too. This entrant: 31 contributing benchmarks, 34406 calibration items; source ten-bin reliability table and area values retained in private evidence. New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported. Uses roughly one sixth of requests with discrete correct answers; benchmarks receive equal weight. ToolRet, BRIGHT and RouterBench are excluded. Known training exposure is penalized.

Checked / reported: 2026-10-20 · Samples: 34406

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Calibration on the source-defined text sample; confidence alone is not a quality ranking.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too. This entrant: 31 contributing benchmarks, 34406 calibration items; source ten-bin reliability table and area values retained in private evidence. New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported. Uses roughly one sixth of requests with discrete correct answers; benchmarks receive equal weight. ToolRet, BRIGHT and RouterBench are excluded. Known training exposure is penalized.

Checked / reported: 2026-10-20 · Samples: 34406

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Calibration on the source-defined text sample; confidence alone is not a quality ranking.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too. This entrant: 31 contributing benchmarks, 34406 calibration items; source ten-bin reliability table and area values retained in private evidence. New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported. Uses roughly one sixth of requests with discrete correct answers; benchmarks receive equal weight. ToolRet, BRIGHT and RouterBench are excluded. Known training exposure is penalized.

Checked / reported: 2026-10-20 · Samples: 34406

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Calibration on the source-defined text sample; confidence alone is not a quality ranking.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too. This entrant: 31 contributing benchmarks, 34406 calibration items; source ten-bin reliability table and area values retained in private evidence. New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported. Uses roughly one sixth of requests with discrete correct answers; benchmarks receive equal weight. ToolRet, BRIGHT and RouterBench are excluded. Known training exposure is penalized.

Checked / reported: 2026-10-20 · Samples: 34406

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Published Median latency for the board setup; hosted network and local inference are separate.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

Local board inference, one RTX PRO 6000; separate from vendor 400-row sample. New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported. Local GPU/runtime can differ between submitted and evaluator runs; not isolated architecture speed.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Published p95 latency for the board setup; hosted network and local inference are separate.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

Local board inference, one RTX PRO 6000; separate from vendor 400-row sample. New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported. Local GPU/runtime can differ between submitted and evaluator runs; not isolated architecture speed.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-reported local mean runtime; hardware/runtime may differ across entrants.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher diagnostic: the share of answers given with at least 95% confidence that were wrong. Preserve publisher denominator and weighting; not an accuracy aggregate.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too. This entrant: 31 contributing benchmarks, 34406 calibration items; source ten-bin reliability table and area values retained in private evidence. New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: 34406

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Source-published category quantity; skill-adjusted, raw and combined values are distinct.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Sample count here is the number of benchmarks in this category, not requests.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: 11

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published aggregate observations; result reuse terms not separately specified

Source-published category quantity; skill-adjusted, raw and combined values are distinct.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Sample count here is the number of benchmarks in this category, not requests.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: 11

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published aggregate observations; result reuse terms not separately specified

Source-published category quantity; skill-adjusted, raw and combined values are distinct.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Sample count here is the number of benchmarks in this category, not requests.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: 11

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published aggregate observations; result reuse terms not separately specified

Source-published category quantity; skill-adjusted, raw and combined values are distinct.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Sample count here is the number of benchmarks in this category, not requests.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: 9

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published aggregate observations; result reuse terms not separately specified

Source-published category quantity; skill-adjusted, raw and combined values are distinct.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Sample count here is the number of benchmarks in this category, not requests.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: 9

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published aggregate observations; result reuse terms not separately specified

Source-published category quantity; skill-adjusted, raw and combined values are distinct.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Sample count here is the number of benchmarks in this category, not requests.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: 9

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published aggregate observations; result reuse terms not separately specified

Source-published category quantity; skill-adjusted, raw and combined values are distinct.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Sample count here is the number of benchmarks in this category, not requests.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: 6

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published aggregate observations; result reuse terms not separately specified

Source-published category quantity; skill-adjusted, raw and combined values are distinct.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Sample count here is the number of benchmarks in this category, not requests.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: 6

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published aggregate observations; result reuse terms not separately specified

Source-published category quantity; skill-adjusted, raw and combined values are distinct.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Sample count here is the number of benchmarks in this category, not requests.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: 6

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published aggregate observations; result reuse terms not separately specified

Source-published category quantity; skill-adjusted, raw and combined values are distinct.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Sample count here is the number of benchmarks in this category, not requests.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: 5

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published aggregate observations; result reuse terms not separately specified

Source-published category quantity; skill-adjusted, raw and combined values are distinct.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Sample count here is the number of benchmarks in this category, not requests.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: 5

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published aggregate observations; result reuse terms not separately specified

Source-published category quantity; skill-adjusted, raw and combined values are distinct.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Sample count here is the number of benchmarks in this category, not requests.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: 5

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published aggregate observations; result reuse terms not separately specified

Source-published category quantity; skill-adjusted, raw and combined values are distinct.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Sample count here is the number of benchmarks in this category, not requests.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: 6

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published aggregate observations; result reuse terms not separately specified

Source-published category quantity; skill-adjusted, raw and combined values are distinct.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Sample count here is the number of benchmarks in this category, not requests.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: 6

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published aggregate observations; result reuse terms not separately specified

Source-published category quantity; skill-adjusted, raw and combined values are distinct.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Sample count here is the number of benchmarks in this category, not requests.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: 6

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published aggregate observations; result reuse terms not separately specified

HF text calibration 0.3.2 · Knowledge · ECE

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too. This entrant: 31 contributing benchmarks, 34406 calibration items; source ten-bin reliability table and area values retained in private evidence. New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Area-specific item count unpublished; do not use whole-sample count.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

HF text calibration 0.3.2 · Knowledge · Sample accuracy

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too. This entrant: 31 contributing benchmarks, 34406 calibration items; source ten-bin reliability table and area values retained in private evidence. New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Area-specific item count unpublished; do not use whole-sample count.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

HF text calibration 0.3.2 · Knowledge · Mean chosen confidence

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too. This entrant: 31 contributing benchmarks, 34406 calibration items; source ten-bin reliability table and area values retained in private evidence. New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Area-specific item count unpublished; do not use whole-sample count.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

HF text calibration 0.3.2 · Language · ECE

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too. This entrant: 31 contributing benchmarks, 34406 calibration items; source ten-bin reliability table and area values retained in private evidence. New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Area-specific item count unpublished; do not use whole-sample count.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

HF text calibration 0.3.2 · Language · Sample accuracy

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too. This entrant: 31 contributing benchmarks, 34406 calibration items; source ten-bin reliability table and area values retained in private evidence. New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Area-specific item count unpublished; do not use whole-sample count.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

HF text calibration 0.3.2 · Language · Mean chosen confidence

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too. This entrant: 31 contributing benchmarks, 34406 calibration items; source ten-bin reliability table and area values retained in private evidence. New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Area-specific item count unpublished; do not use whole-sample count.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

HF text calibration 0.3.2 · Retrieval · ECE

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too. This entrant: 31 contributing benchmarks, 34406 calibration items; source ten-bin reliability table and area values retained in private evidence. New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Area-specific item count unpublished; do not use whole-sample count.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

HF text calibration 0.3.2 · Retrieval · Sample accuracy

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too. This entrant: 31 contributing benchmarks, 34406 calibration items; source ten-bin reliability table and area values retained in private evidence. New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Area-specific item count unpublished; do not use whole-sample count.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

HF text calibration 0.3.2 · Retrieval · Mean chosen confidence

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too. This entrant: 31 contributing benchmarks, 34406 calibration items; source ten-bin reliability table and area values retained in private evidence. New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Area-specific item count unpublished; do not use whole-sample count.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

HF text calibration 0.3.2 · Tools · ECE

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too. This entrant: 31 contributing benchmarks, 34406 calibration items; source ten-bin reliability table and area values retained in private evidence. New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Area-specific item count unpublished; do not use whole-sample count.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

HF text calibration 0.3.2 · Tools · Sample accuracy

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too. This entrant: 31 contributing benchmarks, 34406 calibration items; source ten-bin reliability table and area values retained in private evidence. New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Area-specific item count unpublished; do not use whole-sample count.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

HF text calibration 0.3.2 · Tools · Mean chosen confidence

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too. This entrant: 31 contributing benchmarks, 34406 calibration items; source ten-bin reliability table and area values retained in private evidence. New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Area-specific item count unpublished; do not use whole-sample count.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

HF text calibration 0.3.2 · Arts · ECE

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too. This entrant: 31 contributing benchmarks, 34406 calibration items; source ten-bin reliability table and area values retained in private evidence. New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Area-specific item count unpublished; do not use whole-sample count.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

HF text calibration 0.3.2 · Arts · Sample accuracy

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too. This entrant: 31 contributing benchmarks, 34406 calibration items; source ten-bin reliability table and area values retained in private evidence. New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Area-specific item count unpublished; do not use whole-sample count.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

HF text calibration 0.3.2 · Arts · Mean chosen confidence

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too. This entrant: 31 contributing benchmarks, 34406 calibration items; source ten-bin reliability table and area values retained in private evidence. New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Area-specific item count unpublished; do not use whole-sample count.

32 benchmarks with a per-field right or wrong answer (ToolRet, BRIGHT and RouterBench excluded), each weighted equally; a 1-in-6 sample of requests; confidence is the probability placed on the chosen option. Rows an entrant trained on count as wrong here too.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

BFCL formulated as typed decisions; published case exact accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "case exact accuracy", "requests": 1694, "answered": 1694, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 1694

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native case exact accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9776, "skill": 0.9697, "coverage": 1, "random": 0.2592, "tracks": [], "combined": 0.9434, "private_eq": 0.9501}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 1694

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native case exact accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9776, "skill": 0.9697, "coverage": 1, "random": 0.2592, "tracks": [], "combined": 0.9434, "private_eq": 0.9501}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 1694

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native case exact accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9776, "skill": 0.9697, "coverage": 1, "random": 0.2592, "tracks": [], "combined": 0.9434, "private_eq": 0.9501}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native case exact accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9776, "skill": 0.9697, "coverage": 1, "random": 0.2592, "tracks": [], "combined": 0.9434, "private_eq": 0.9501}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

ToolRet formulated as typed decisions; published nDCG@10 for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "nDCG@10", "requests": 685, "answered": 685, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 685

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native nDCG@10.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.6752, "skill": 0.6249, "coverage": 1, "random": 0.1341, "tracks": [], "combined": 0.6799, "private_eq": 0.744}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 685

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native nDCG@10.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.6752, "skill": 0.6249, "coverage": 1, "random": 0.1341, "tracks": [], "combined": 0.6799, "private_eq": 0.744}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 685

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native nDCG@10.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.6752, "skill": 0.6249, "coverage": 1, "random": 0.1341, "tracks": [], "combined": 0.6799, "private_eq": 0.744}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native nDCG@10.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.6752, "skill": 0.6249, "coverage": 1, "random": 0.1341, "tracks": [], "combined": 0.6799, "private_eq": 0.744}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

API-Bank formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy", "requests": 508, "answered": 508, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 508

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.874, "skill": 0.8716, "coverage": 1, "random": 0.0189, "tracks": [], "combined": 0.9116, "private_eq": 0.9288}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 508

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.874, "skill": 0.8716, "coverage": 1, "random": 0.0189, "tracks": [], "combined": 0.9116, "private_eq": 0.9288}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 508

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.874, "skill": 0.8716, "coverage": 1, "random": 0.0189, "tracks": [], "combined": 0.9116, "private_eq": 0.9288}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.874, "skill": 0.8716, "coverage": 1, "random": 0.0189, "tracks": [], "combined": 0.9116, "private_eq": 0.9288}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

BANKING77 formulated as typed decisions; published macro-F1 for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "macro-F1", "requests": 3080, "answered": 3080, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 3080

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8565, "skill": 0.8547, "coverage": 1, "random": 0.0127, "tracks": [], "combined": 0.8745, "private_eq": 0.8842}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 3080

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8565, "skill": 0.8547, "coverage": 1, "random": 0.0127, "tracks": [], "combined": 0.8745, "private_eq": 0.8842}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 3080

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8565, "skill": 0.8547, "coverage": 1, "random": 0.0127, "tracks": [], "combined": 0.8745, "private_eq": 0.8842}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8565, "skill": 0.8547, "coverage": 1, "random": 0.0127, "tracks": [], "combined": 0.8745, "private_eq": 0.8842}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

CLINC150+OOS formulated as typed decisions; published macro-F1 for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "macro-F1", "requests": 5500, "answered": 5500, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 5500

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9355, "skill": 0.9351, "coverage": 1, "random": 0.006, "tracks": [], "combined": 0.8044, "private_eq": 0.7554}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 5500

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9355, "skill": 0.9351, "coverage": 1, "random": 0.006, "tracks": [], "combined": 0.8044, "private_eq": 0.7554}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 5500

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9355, "skill": 0.9351, "coverage": 1, "random": 0.006, "tracks": [], "combined": 0.8044, "private_eq": 0.7554}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9355, "skill": 0.9351, "coverage": 1, "random": 0.006, "tracks": [], "combined": 0.8044, "private_eq": 0.7554}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

RouterBench formulated as typed decisions; published selected quality (quality objective) for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "selected quality (quality objective)", "requests": 10000, "answered": 10000, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 10000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native selected quality (quality objective).

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.7998, "skill": 0.5306, "coverage": 1, "random": 0.5735, "tracks": []}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 10000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native selected quality (quality objective).

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.7998, "skill": 0.5306, "coverage": 1, "random": 0.5735, "tracks": []}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 10000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Home appliance simulator formulated as typed decisions; published case exact accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "case exact accuracy", "requests": 88, "answered": 88, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 88

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native case exact accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.625, "skill": 0.625, "coverage": 1, "random": 0, "tracks": [], "combined": 0.6425, "private_eq": 0.6525}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 88

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native case exact accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.625, "skill": 0.625, "coverage": 1, "random": 0, "tracks": [], "combined": 0.6425, "private_eq": 0.6525}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 88

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native case exact accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.625, "skill": 0.625, "coverage": 1, "random": 0, "tracks": [], "combined": 0.6425, "private_eq": 0.6525}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native case exact accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.625, "skill": 0.625, "coverage": 1, "random": 0, "tracks": [], "combined": 0.6425, "private_eq": 0.6525}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

SGD/SGD-X formulated as typed decisions; published macro-F1 for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "macro-F1", "requests": 2500, "answered": 2500, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 2500

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.3219, "skill": 0, "coverage": 1, "random": 0.399, "tracks": []}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 2500

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.3219, "skill": 0, "coverage": 1, "random": 0.399, "tracks": []}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 2500

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

ContractNLI formulated as typed decisions; published macro-F1 for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "macro-F1", "requests": 123, "answered": 123, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 123

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8125, "skill": 0.7289, "coverage": 1, "random": 0.3085, "tracks": []}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 123

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8125, "skill": 0.7289, "coverage": 1, "random": 0.3085, "tracks": []}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 123

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

ANLI formulated as typed decisions; published macro-F1 for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "macro-F1", "requests": 3200, "answered": 3200, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 3200

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.7686, "skill": 0.6534, "coverage": 1, "random": 0.3324, "tracks": [], "combined": 0.5777, "private_eq": 0.7006}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 3200

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.7686, "skill": 0.6534, "coverage": 1, "random": 0.3324, "tracks": [], "combined": 0.5777, "private_eq": 0.7006}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 3200

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.7686, "skill": 0.6534, "coverage": 1, "random": 0.3324, "tracks": [], "combined": 0.5777, "private_eq": 0.7006}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.7686, "skill": 0.6534, "coverage": 1, "random": 0.3324, "tracks": [], "combined": 0.5777, "private_eq": 0.7006}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

BPoMP formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy", "requests": 5000, "answered": 5000, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 5000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9251, "skill": 0.8502, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.8753, "private_eq": 0.9428}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 5000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9251, "skill": 0.8502, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.8753, "private_eq": 0.9428}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 5000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9251, "skill": 0.8502, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.8753, "private_eq": 0.9428}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9251, "skill": 0.8502, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.8753, "private_eq": 0.9428}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Humicroedit formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy", "requests": 2628, "answered": 2628, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 2628

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.6416, "skill": 0.2831, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.2713, "private_eq": 0.637}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 2628

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.6416, "skill": 0.2831, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.2713, "private_eq": 0.637}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 2628

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.6416, "skill": 0.2831, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.2713, "private_eq": 0.637}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.6416, "skill": 0.2831, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.2713, "private_eq": 0.637}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

POP909-CL formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy", "requests": 2000, "answered": 2000, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 2000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.2671, "skill": 0.2613, "coverage": 1, "random": 0.0078, "tracks": [], "combined": 0.2592, "private_eq": 0.2715}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 2000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.2671, "skill": 0.2613, "coverage": 1, "random": 0.0078, "tracks": [], "combined": 0.2592, "private_eq": 0.2715}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 2000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.2671, "skill": 0.2613, "coverage": 1, "random": 0.0078, "tracks": [], "combined": 0.2592, "private_eq": 0.2715}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.2671, "skill": 0.2613, "coverage": 1, "random": 0.0078, "tracks": [], "combined": 0.2592, "private_eq": 0.2715}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

cfcolor formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy", "requests": 5000, "answered": 5000, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 5000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.6527, "skill": 0.3053, "coverage": 1, "random": 0.5, "tracks": []}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 5000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.6527, "skill": 0.3053, "coverage": 1, "random": 0.5, "tracks": []}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 5000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

MMLU formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy", "requests": 14033, "answered": 14033, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 14033

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8686, "skill": 0.8248, "coverage": 1, "random": 0.25, "tracks": []}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 14033

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8686, "skill": 0.8248, "coverage": 1, "random": 0.25, "tracks": []}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 14033

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

GPQA Diamond formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy", "requests": 196, "answered": 196, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 196

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.4949, "skill": 0.3265, "coverage": 1, "random": 0.25, "tracks": [], "private_eq": 0.521, "combined": 0.3468}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 196

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.4949, "skill": 0.3265, "coverage": 1, "random": 0.25, "tracks": [], "private_eq": 0.521, "combined": 0.3468}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 196

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.4949, "skill": 0.3265, "coverage": 1, "random": 0.25, "tracks": [], "private_eq": 0.521, "combined": 0.3468}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.4949, "skill": 0.3265, "coverage": 1, "random": 0.25, "tracks": [], "private_eq": 0.521, "combined": 0.3468}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

ARC-Easy formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy", "requests": 2376, "answered": 2376, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 2376

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9916, "skill": 0.9888, "coverage": 1, "random": 0.2502, "tracks": []}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 2376

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9916, "skill": 0.9888, "coverage": 1, "random": 0.2502, "tracks": []}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 2376

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

ARC-Challenge formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy", "requests": 1172, "answered": 1172, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 1172

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9744, "skill": 0.9659, "coverage": 1, "random": 0.2502, "tracks": []}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 1172

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9744, "skill": 0.9659, "coverage": 1, "random": 0.2502, "tracks": []}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 1172

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

WinoGrande formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy", "requests": 1267, "answered": 1267, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 1267

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9234, "skill": 0.8468, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.7012, "private_eq": 0.823}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 1267

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9234, "skill": 0.8468, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.7012, "private_eq": 0.823}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 1267

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9234, "skill": 0.8468, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.7012, "private_eq": 0.823}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9234, "skill": 0.8468, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.7012, "private_eq": 0.823}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

HellaSwag formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy", "requests": 10042, "answered": 10042, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 10042

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9618, "skill": 0.9491, "coverage": 1, "random": 0.25, "tracks": [], "combined": 0.9862, "private_eq": 1}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 10042

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9618, "skill": 0.9491, "coverage": 1, "random": 0.25, "tracks": [], "combined": 0.9862, "private_eq": 1}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 10042

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9618, "skill": 0.9491, "coverage": 1, "random": 0.25, "tracks": [], "combined": 0.9862, "private_eq": 1}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9618, "skill": 0.9491, "coverage": 1, "random": 0.25, "tracks": [], "combined": 0.9862, "private_eq": 1}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

GSM8K formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy (4- and 10-choice)", "requests": 2638, "answered": 2638, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 2638

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8624, "skill": 0.8346, "coverage": 1, "random": 0.175, "tracks": [{"track": "4-choice", "raw": 0.8878, "random": 0.25}, {"track": "10-choice", "raw": 0.837, "random": 0.1}], "combined": 0.8126, "private_eq": 0.8391}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 2638

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8624, "skill": 0.8346, "coverage": 1, "random": 0.175, "tracks": [{"track": "4-choice", "raw": 0.8878, "random": 0.25}, {"track": "10-choice", "raw": 0.837, "random": 0.1}], "combined": 0.8126, "private_eq": 0.8391}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 2638

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8624, "skill": 0.8346, "coverage": 1, "random": 0.175, "tracks": [{"track": "4-choice", "raw": 0.8878, "random": 0.25}, {"track": "10-choice", "raw": 0.837, "random": 0.1}], "combined": 0.8126, "private_eq": 0.8391}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8624, "skill": 0.8346, "coverage": 1, "random": 0.175, "tracks": [{"track": "4-choice", "raw": 0.8878, "random": 0.25}, {"track": "10-choice", "raw": 0.837, "random": 0.1}], "combined": 0.8126, "private_eq": 0.8391}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Separate publisher subtask/track; track sample count is not stated in the display.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Published track {"track": "4-choice", "raw": 0.8878, "random": 0.25}

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Separate publisher subtask/track; track sample count is not stated in the display.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Published track {"track": "10-choice", "raw": 0.837, "random": 0.1}

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

ChessBench formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy", "requests": 5000, "answered": 5000, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 5000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.236, "skill": 0.1679, "coverage": 1, "random": 0.0819, "tracks": [], "combined": 0.1089, "private_eq": 0.1689}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 5000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.236, "skill": 0.1679, "coverage": 1, "random": 0.0819, "tracks": [], "combined": 0.1089, "private_eq": 0.1689}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 5000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.236, "skill": 0.1679, "coverage": 1, "random": 0.0819, "tracks": [], "combined": 0.1089, "private_eq": 0.1689}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.236, "skill": 0.1679, "coverage": 1, "random": 0.0819, "tracks": [], "combined": 0.1089, "private_eq": 0.1689}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

MuSR formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy", "requests": 752, "answered": 752, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 752

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.6503, "skill": 0.444, "coverage": 1, "random": 0.371, "tracks": [], "combined": 0.3665, "private_eq": 0.5862}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 752

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.6503, "skill": 0.444, "coverage": 1, "random": 0.371, "tracks": [], "combined": 0.3665, "private_eq": 0.5862}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 752

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.6503, "skill": 0.444, "coverage": 1, "random": 0.371, "tracks": [], "combined": 0.3665, "private_eq": 0.5862}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.6503, "skill": 0.444, "coverage": 1, "random": 0.371, "tracks": [], "combined": 0.3665, "private_eq": 0.5862}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

SATA-Bench formulated as typed decisions; published case exact accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "case exact accuracy", "requests": 1650, "answered": 1650, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 1650

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native case exact accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.1145, "skill": 0.1027, "coverage": 1, "random": 0.0131, "tracks": [], "combined": 0.0764, "private_eq": 0.0877}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 1650

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native case exact accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.1145, "skill": 0.1027, "coverage": 1, "random": 0.0131, "tracks": [], "combined": 0.0764, "private_eq": 0.0877}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 1650

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native case exact accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.1145, "skill": 0.1027, "coverage": 1, "random": 0.0131, "tracks": [], "combined": 0.0764, "private_eq": 0.0877}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native case exact accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.1145, "skill": 0.1027, "coverage": 1, "random": 0.0131, "tracks": [], "combined": 0.0764, "private_eq": 0.0877}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

BRIGHT formulated as typed decisions; published nDCG@10 for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "nDCG@10", "requests": 220, "answered": 220, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 220

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native nDCG@10.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.4617, "skill": 0.391, "coverage": 1, "random": 0.116, "tracks": [], "combined": 0.4087, "private_eq": 0.4886}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 220

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native nDCG@10.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.4617, "skill": 0.391, "coverage": 1, "random": 0.116, "tracks": [], "combined": 0.4087, "private_eq": 0.4886}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 220

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native nDCG@10.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.4617, "skill": 0.391, "coverage": 1, "random": 0.116, "tracks": [], "combined": 0.4087, "private_eq": 0.4886}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native nDCG@10.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.4617, "skill": 0.391, "coverage": 1, "random": 0.116, "tracks": [], "combined": 0.4087, "private_eq": 0.4886}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Amazon ESCI formulated as typed decisions; published macro-F1 for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "macro-F1", "requests": 5000, "answered": 5000, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 5000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.5731, "skill": 0.4645, "coverage": 1, "random": 0.2027, "tracks": [], "combined": 0.5343, "private_eq": 0.6541}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 5000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.5731, "skill": 0.4645, "coverage": 1, "random": 0.2027, "tracks": [], "combined": 0.5343, "private_eq": 0.6541}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 5000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.5731, "skill": 0.4645, "coverage": 1, "random": 0.2027, "tracks": [], "combined": 0.5343, "private_eq": 0.6541}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.5731, "skill": 0.4645, "coverage": 1, "random": 0.2027, "tracks": [], "combined": 0.5343, "private_eq": 0.6541}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

ACOS formulated as typed decisions; published per-review F1 for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "per-review F1", "requests": 1565, "answered": 1565, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 1565

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native per-review F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.2641, "skill": 0.2406, "coverage": 1, "random": 0.031, "tracks": [], "combined": 0.3749, "private_eq": 0.4517}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 1565

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native per-review F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.2641, "skill": 0.2406, "coverage": 1, "random": 0.031, "tracks": [], "combined": 0.3749, "private_eq": 0.4517}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 1565

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native per-review F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.2641, "skill": 0.2406, "coverage": 1, "random": 0.031, "tracks": [], "combined": 0.3749, "private_eq": 0.4517}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native per-review F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.2641, "skill": 0.2406, "coverage": 1, "random": 0.031, "tracks": [], "combined": 0.3749, "private_eq": 0.4517}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

FinEntity formulated as typed decisions; published macro-F1 for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "macro-F1", "requests": 979, "answered": 979, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 979

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9292, "skill": 0.8959, "coverage": 1, "random": 0.3201, "tracks": [], "combined": 0.595, "private_eq": 0.6462}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 979

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9292, "skill": 0.8959, "coverage": 1, "random": 0.3201, "tracks": [], "combined": 0.595, "private_eq": 0.6462}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 979

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9292, "skill": 0.8959, "coverage": 1, "random": 0.3201, "tracks": [], "combined": 0.595, "private_eq": 0.6462}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9292, "skill": 0.8959, "coverage": 1, "random": 0.3201, "tracks": [], "combined": 0.595, "private_eq": 0.6462}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

iSarcasmEval formulated as typed decisions; published Sarcasm F1 · track A, English for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "Sarcasm F1 \u00b7 track A, English", "requests": 4600, "answered": 4600, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 4600

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native Sarcasm F1 · track A, English.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.6189, "skill": 0.5097, "coverage": 1, "random": 0.2227, "tracks": [{"track": "A · Arabic", "score": 0.6054, "headline": false}, {"track": "A · English", "score": 0.6189, "headline": true}, {"track": "C · Arabic pairs", "score": 0.975, "headline": false}, {"track": "C · English pairs", "score": 0.99, "headline": false}], "combined": 0.5474, "private_eq": 0.663}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 4600

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native Sarcasm F1 · track A, English.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.6189, "skill": 0.5097, "coverage": 1, "random": 0.2227, "tracks": [{"track": "A · Arabic", "score": 0.6054, "headline": false}, {"track": "A · English", "score": 0.6189, "headline": true}, {"track": "C · Arabic pairs", "score": 0.975, "headline": false}, {"track": "C · English pairs", "score": 0.99, "headline": false}], "combined": 0.5474, "private_eq": 0.663}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 4600

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native Sarcasm F1 · track A, English.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.6189, "skill": 0.5097, "coverage": 1, "random": 0.2227, "tracks": [{"track": "A · Arabic", "score": 0.6054, "headline": false}, {"track": "A · English", "score": 0.6189, "headline": true}, {"track": "C · Arabic pairs", "score": 0.975, "headline": false}, {"track": "C · English pairs", "score": 0.99, "headline": false}], "combined": 0.5474, "private_eq": 0.663}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native Sarcasm F1 · track A, English.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.6189, "skill": 0.5097, "coverage": 1, "random": 0.2227, "tracks": [{"track": "A · Arabic", "score": 0.6054, "headline": false}, {"track": "A · English", "score": 0.6189, "headline": true}, {"track": "C · Arabic pairs", "score": 0.975, "headline": false}, {"track": "C · English pairs", "score": 0.99, "headline": false}], "combined": 0.5474, "private_eq": 0.663}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Separate publisher subtask/track; track sample count is not stated in the display.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Published track {"track": "A · Arabic", "score": 0.6054, "headline": false}

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Separate publisher subtask/track; track sample count is not stated in the display.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Published track {"track": "C · Arabic pairs", "score": 0.975, "headline": false}

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Separate publisher subtask/track; track sample count is not stated in the display.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Published track {"track": "C · English pairs", "score": 0.99, "headline": false}

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

VAST formulated as typed decisions; published macro-F1 for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "macro-F1", "requests": 3006, "answered": 3006, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 3006

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.7315, "skill": 0.5973, "coverage": 1, "random": 0.3333, "tracks": [], "combined": 0.5564, "private_eq": 0.6962}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 3006

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.7315, "skill": 0.5973, "coverage": 1, "random": 0.3333, "tracks": [], "combined": 0.5564, "private_eq": 0.6962}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 3006

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.7315, "skill": 0.5973, "coverage": 1, "random": 0.3333, "tracks": [], "combined": 0.5564, "private_eq": 0.6962}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.7315, "skill": 0.5973, "coverage": 1, "random": 0.3333, "tracks": [], "combined": 0.5564, "private_eq": 0.6962}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

NLI4CT formulated as typed decisions; published macro-F1 for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "macro-F1", "requests": 5500, "answered": 5500, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 5500

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8441, "skill": 0.6967, "coverage": 1, "random": 0.486, "tracks": [], "combined": 0.8455, "private_eq": 0.9512}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 5500

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8441, "skill": 0.6967, "coverage": 1, "random": 0.486, "tracks": [], "combined": 0.8455, "private_eq": 0.9512}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 5500

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8441, "skill": 0.6967, "coverage": 1, "random": 0.486, "tracks": [], "combined": 0.8455, "private_eq": 0.9512}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native macro-F1.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8441, "skill": 0.6967, "coverage": 1, "random": 0.486, "tracks": [], "combined": 0.8455, "private_eq": 0.9512}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

CRUXEval formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy", "requests": 570, "answered": 570, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 570

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8386, "skill": 0.7439, "coverage": 1, "random": 0.3697, "tracks": [], "combined": 0.6887, "private_eq": 0.7916}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 570

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8386, "skill": 0.7439, "coverage": 1, "random": 0.3697, "tracks": [], "combined": 0.6887, "private_eq": 0.7916}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 570

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8386, "skill": 0.7439, "coverage": 1, "random": 0.3697, "tracks": [], "combined": 0.6887, "private_eq": 0.7916}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8386, "skill": 0.7439, "coverage": 1, "random": 0.3697, "tracks": [], "combined": 0.6887, "private_eq": 0.7916}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

CLadder formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy", "requests": 5000, "answered": 5000, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 5000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8016, "skill": 0.6032, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.6005, "private_eq": 0.8015}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 5000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8016, "skill": 0.6032, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.6005, "private_eq": 0.8015}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 5000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8016, "skill": 0.6032, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.6005, "private_eq": 0.8015}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8016, "skill": 0.6032, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.6005, "private_eq": 0.8015}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

HLE formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy", "requests": 501, "answered": 501, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 501

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.1018, "skill": 0, "coverage": 1, "random": 0.1641, "tracks": []}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 501

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.1018, "skill": 0, "coverage": 1, "random": 0.1641, "tracks": []}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 501

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Habermas Machine formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy", "requests": 1676, "answered": 1676, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 1676

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.4433, "skill": 0.1921, "coverage": 1, "random": 0.311, "tracks": [], "combined": 0.2155, "private_eq": 0.4713}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 1676

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.4433, "skill": 0.1921, "coverage": 1, "random": 0.311, "tracks": [], "combined": 0.2155, "private_eq": 0.4713}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 1676

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.4433, "skill": 0.1921, "coverage": 1, "random": 0.311, "tracks": [], "combined": 0.2155, "private_eq": 0.4713}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.4433, "skill": 0.1921, "coverage": 1, "random": 0.311, "tracks": [], "combined": 0.2155, "private_eq": 0.4713}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

PhishNChips phishing decisions formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy", "requests": 2000, "answered": 2000, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 2000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9245, "skill": 0.849, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.8403, "private_eq": 0.9188}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 2000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9245, "skill": 0.849, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.8403, "private_eq": 0.9188}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 2000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9245, "skill": 0.849, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.8403, "private_eq": 0.9188}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.9245, "skill": 0.849, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.8403, "private_eq": 0.9188}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

MMLU-Pro formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy", "requests": 12032, "answered": 12032, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 12032

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.7197, "skill": 0.6847, "coverage": 1, "random": 0.1109, "tracks": [], "private_eq": 0.5352, "combined": 0.5329}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 12032

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.7197, "skill": 0.6847, "coverage": 1, "random": 0.1109, "tracks": [], "private_eq": 0.5352, "combined": 0.5329}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 12032

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.7197, "skill": 0.6847, "coverage": 1, "random": 0.1109, "tracks": [], "private_eq": 0.5352, "combined": 0.5329}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.7197, "skill": 0.6847, "coverage": 1, "random": 0.1109, "tracks": [], "private_eq": 0.5352, "combined": 0.5329}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

BBH formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy", "requests": 5507, "answered": 5507, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 5507

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8064, "skill": 0.7194, "coverage": 1, "random": 0.3101, "tracks": [], "private_eq": 0.5354, "combined": 0.434}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 5507

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8064, "skill": 0.7194, "coverage": 1, "random": 0.3101, "tracks": [], "private_eq": 0.5354, "combined": 0.434}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 5507

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8064, "skill": 0.7194, "coverage": 1, "random": 0.3101, "tracks": [], "private_eq": 0.5354, "combined": 0.434}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8064, "skill": 0.7194, "coverage": 1, "random": 0.3101, "tracks": [], "private_eq": 0.5354, "combined": 0.434}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

RAGTruth formulated as typed decisions; published F1 on hallucinated class for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "F1 on hallucinated class", "requests": 2700, "answered": 2700, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 2700

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native F1 on hallucinated class.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.7944, "skill": 0.5737, "coverage": 1, "random": 0.5177, "tracks": [], "combined": 0.4734, "private_eq": 0.7293}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 2700

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native F1 on hallucinated class.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.7944, "skill": 0.5737, "coverage": 1, "random": 0.5177, "tracks": [], "combined": 0.4734, "private_eq": 0.7293}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 2700

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native F1 on hallucinated class.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.7944, "skill": 0.5737, "coverage": 1, "random": 0.5177, "tracks": [], "combined": 0.4734, "private_eq": 0.7293}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native F1 on hallucinated class.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.7944, "skill": 0.5737, "coverage": 1, "random": 0.5177, "tracks": [], "combined": 0.4734, "private_eq": 0.7293}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

HoVer formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy", "requests": 4000, "answered": 4000, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 4000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.83, "skill": 0.66, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.6848, "private_eq": 0.8486}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 4000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.83, "skill": 0.66, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.6848, "private_eq": 0.8486}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 4000

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.83, "skill": 0.66, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.6848, "private_eq": 0.8486}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.83, "skill": 0.66, "coverage": 1, "random": 0.5, "tracks": [], "combined": 0.6848, "private_eq": 0.8486}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

When2Call formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy", "requests": 3652, "answered": 3652, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 3652

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8436, "skill": 0.7915, "coverage": 1, "random": 0.25, "tracks": [], "combined": 0.6332, "private_eq": 0.6803}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 3652

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8436, "skill": 0.7915, "coverage": 1, "random": 0.25, "tracks": [], "combined": 0.6332, "private_eq": 0.6803}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 3652

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8436, "skill": 0.7915, "coverage": 1, "random": 0.25, "tracks": [], "combined": 0.6332, "private_eq": 0.6803}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.8436, "skill": 0.7915, "coverage": 1, "random": 0.25, "tracks": [], "combined": 0.6332, "private_eq": 0.6803}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

New Yorker formulated as typed decisions; published accuracy for this board population.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Native board observation; support/status: {"metric": "accuracy", "requests": 528, "answered": 528, "unsupported": 0, "errors": 0}. Per-benchmark zero timings are batch placeholders, not measured inference speed.

Community Decision Index 0.3.2; distinct from AA. Full = 20% public + 50% private same-skills + 30% private new-domains on publisher-normalized scales. Public suite is chance-corrected and coverage-adjusted; unanswered counts wrong. Some public runs are author-submitted and private runs evaluator-run; each record preserves that status. Small differences can be tied by the publisher. Snapshot date is not execution date. No underlying private test text is imported.

Checked / reported: 2026-10-20 · Samples: 528

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined raw index component; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.7746, "skill": 0.7182, "coverage": 1, "random": 0.2, "tracks": [], "combined": 0.747, "private_eq": 0.8082}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 528

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined chance-corrected public skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.7746, "skill": 0.7182, "coverage": 1, "random": 0.2, "tracks": [], "combined": 0.747, "private_eq": 0.8082}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: 528

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined combined public/private; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.7746, "skill": 0.7182, "coverage": 1, "random": 0.2, "tracks": [], "combined": 0.747, "private_eq": 0.8082}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-defined equated private same-skill; source normalized units distinct from native accuracy.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z. Raw index component can use a balanced/cluster-macro convention distinct from the ordinary native result. Published normalization context: {"raw": 0.7746, "skill": 0.7182, "coverage": 1, "random": 0.2, "tracks": [], "combined": 0.747, "private_eq": 0.8082}

Published quantity; no recomputation from rounded rows. Source chance/coverage and tracks retained in notes.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-equated private component on public scale; separate from un-equated private component.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-equated private component on public scale; separate from un-equated private component.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-reported weighted contribution to the Full score; not a separate benchmark execution.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-reported weighted contribution to the Full score; not a separate benchmark execution.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Publisher-reported weighted contribution to the Full score; not a separate benchmark execution.

Exact sieve-31b / sthanika-ai/Sieve-31B. Author-submitted Decision Index 0.3 public run adopted after kit rescore; board-run private tests, RTX PRO 6000, author serving code. Public execution 2026-10-10; private execution date unpublished. Adapter unmerged for 42% of private rows, merged for remainder; near-tie differences disclosed. BF16 base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475; adapter SHA256 ea55d2ff60d484dababb9d4042333667cb9d2e59436044021daf4ea2894807bd; T=8.5495. Not a fully independently executed public+private run.

New in the 2026-10-12 PR batch (PR #149). Submitted public 0.3 run adopted (kit rescore 62.52, equal to the claim). Private tests run by the lab on one RTX PRO 6000 with the author's serving; decision agreement with the submitted rows 99.53% on 827 rows (bar 98.5%). Training: about 190k questions from public training splits plus programmatic and synthetic tasks; the author reports a near-duplicate check against every suite row, no per-dataset list. Submitted rows came from the author's batched evaluator, re-tempered offline; the card ran the published engine one request at a time, so identity rests on the agreement gate. The commit the run loaded was removed by a history squash; the adapter weights are identical in the surviving commit. Private run: 42% of rows with the adapter unmerged as published, the rest with it merged into the base for speed (spot check: one exact-tie flip in 258 questions). Small Full-score differences can be publisher ties. Raw rank fields are not a universal current position; no second-place claim. Full is 20% public plus 50% equated private same-skills plus 30% equated private new domains. Private normalized observations are not public accuracy. Source snapshot generated 2026-10-16T16:39:08Z.

Checked / reported: 2026-10-20 · Samples: Not published

Snapshot: 2026-10-16T16:39:08Z · Revision: 0ddf6758073eb77ae449a3541f57286d7f39d93f · Published observations; result reuse terms not specified

Sieve vendor DI 0.3 · Public index

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 140178

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · Raw index

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 140178

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · Public breadth index

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 140178

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · Knowledge & Reasoning · Raw index

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Sample count is contributing benchmarks, not request count.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 11

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · Knowledge & Reasoning · Public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Sample count is contributing benchmarks, not request count.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 11

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · Language Understanding · Raw index

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Sample count is contributing benchmarks, not request count.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 9

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · Language Understanding · Public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Sample count is contributing benchmarks, not request count.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 9

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · Retrieval & Classification · Raw index

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Sample count is contributing benchmarks, not request count.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 6

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · Retrieval & Classification · Public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Sample count is contributing benchmarks, not request count.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 6

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · Tools & Automation · Raw index

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Sample count is contributing benchmarks, not request count.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · Tools & Automation · Public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Sample count is contributing benchmarks, not request count.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · Arts & Human Taste · Raw index

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Sample count is contributing benchmarks, not request count.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 6

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · Arts & Human Taste · Public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Sample count is contributing benchmarks, not request count.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 6

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. case exact accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 1, "dataset": "BFCL", "requests": 1694, "answered": 1694, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 1694, "metric": "case exact accuracy", "reference_same_cases": null, "index_raw": 0.9776, "index_skill": 0.9697, "coverage": 1.0, "chance": 0.2592, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1694

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · BFCL · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.9776, "skill": 0.9697, "coverage": 1.0, "random": 0.2592, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1694

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · BFCL · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.9776, "skill": 0.9697, "coverage": 1.0, "random": 0.2592, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1694

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. nDCG@10

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 2, "dataset": "ToolRet", "requests": 685, "answered": 685, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 685, "metric": "nDCG@10", "reference_same_cases": null, "index_raw": 0.6752, "index_skill": 0.6249, "coverage": 1.0, "chance": 0.1341, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 685

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · ToolRet · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.6752, "skill": 0.6249, "coverage": 1.0, "random": 0.1341, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 685

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · ToolRet · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.6752, "skill": 0.6249, "coverage": 1.0, "random": 0.1341, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 685

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 3, "dataset": "API-Bank", "requests": 508, "answered": 508, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 508, "metric": "accuracy", "reference_same_cases": null, "index_raw": 0.874, "index_skill": 0.8716, "coverage": 1.0, "chance": 0.0189, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 508

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · API-Bank · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.874, "skill": 0.8716, "coverage": 1.0, "random": 0.0189, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 508

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · API-Bank · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.874, "skill": 0.8716, "coverage": 1.0, "random": 0.0189, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 508

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. macro-F1

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 4, "dataset": "BANKING77", "requests": 3080, "answered": 3080, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 3080, "metric": "macro-F1", "reference_same_cases": null, "index_raw": 0.8565, "index_skill": 0.8547, "coverage": 1.0, "chance": 0.0127, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 3080

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · BANKING77 · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.8565, "skill": 0.8547, "coverage": 1.0, "random": 0.0127, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 3080

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · BANKING77 · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.8565, "skill": 0.8547, "coverage": 1.0, "random": 0.0127, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 3080

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. macro-F1

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 5, "dataset": "CLINC150+OOS", "requests": 5500, "answered": 5500, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 5500, "metric": "macro-F1", "reference_same_cases": null, "index_raw": 0.9355, "index_skill": 0.9351, "coverage": 1.0, "chance": 0.006, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5500

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · CLINC150+OOS · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.9355, "skill": 0.9351, "coverage": 1.0, "random": 0.006, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5500

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · CLINC150+OOS · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.9355, "skill": 0.9351, "coverage": 1.0, "random": 0.006, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5500

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. selected quality (quality objective)

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 6, "dataset": "RouterBench", "requests": 10000, "answered": 10000, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 10000, "metric": "selected quality (quality objective)", "reference_same_cases": null, "index_raw": 0.7998, "index_skill": 0.5306, "coverage": 1.0, "chance": 0.5735, "in_index": false}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 10000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · RouterBench · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.7998, "skill": 0.5306, "coverage": 1.0, "random": 0.5735, "rule": "shown, not counted", "in_index": false}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 10000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · RouterBench · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.7998, "skill": 0.5306, "coverage": 1.0, "random": 0.5735, "rule": "shown, not counted", "in_index": false}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 10000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · RouterBench · RouterBench-0shot

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Separate subtask/track; source track support {"track": "RouterBench-0shot", "metric": "selected quality (quality objective)", "score": 0.7862782330601639, "scored_requests": 5003}.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5003

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · RouterBench · RouterBench-5shot

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Separate subtask/track; source track support {"track": "RouterBench-5shot", "metric": "selected quality (quality objective)", "score": 0.8132879727836702, "scored_requests": 4997}.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 4997

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. case exact accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 9, "dataset": "Home appliance simulator", "requests": 88, "answered": 88, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 88, "metric": "case exact accuracy", "reference_same_cases": null, "index_raw": 0.625, "index_skill": 0.625, "coverage": 1.0, "chance": 0.0, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 88

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · Home appliance simulator · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.625, "skill": 0.625, "coverage": 1.0, "random": 0.0, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 88

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · Home appliance simulator · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.625, "skill": 0.625, "coverage": 1.0, "random": 0.0, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 88

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. macro-F1

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 10, "dataset": "SGD/SGD-X", "requests": 2500, "answered": 2500, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 2500, "metric": "macro-F1", "reference_same_cases": null, "index_raw": 0.3219, "index_skill": 0.0, "coverage": 1.0, "chance": 0.399, "in_index": false}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2500

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · SGD/SGD-X · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.3219, "skill": 0.0, "coverage": 1.0, "random": 0.399, "rule": "shown, not counted", "in_index": false}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2500

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · SGD/SGD-X · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.3219, "skill": 0.0, "coverage": 1.0, "random": 0.399, "rule": "shown, not counted", "in_index": false}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2500

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. macro-F1

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 11, "dataset": "ContractNLI", "requests": 123, "answered": 123, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 123, "metric": "macro-F1", "reference_same_cases": null, "index_raw": 0.8125, "index_skill": 0.7289, "coverage": 1.0, "chance": 0.3085, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 123

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · ContractNLI · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.8125, "skill": 0.7289, "coverage": 1.0, "random": 0.3085, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 123

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · ContractNLI · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.8125, "skill": 0.7289, "coverage": 1.0, "random": 0.3085, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 123

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. macro-F1

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 12, "dataset": "ANLI", "requests": 3200, "answered": 3200, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 3200, "metric": "macro-F1", "reference_same_cases": null, "index_raw": 0.7686, "index_skill": 0.6534, "coverage": 1.0, "chance": 0.3324, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 3200

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · ANLI · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.7686, "skill": 0.6534, "coverage": 1.0, "random": 0.3324, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 3200

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · ANLI · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.7686, "skill": 0.6534, "coverage": 1.0, "random": 0.3324, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 3200

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 20, "dataset": "BPoMP", "requests": 5000, "answered": 5000, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 5000, "metric": "accuracy", "reference_same_cases": null, "index_raw": 0.9251, "index_skill": 0.8502, "coverage": 1.0, "chance": 0.5, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · BPoMP · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.9251, "skill": 0.8502, "coverage": 1.0, "random": 0.5, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · BPoMP · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.9251, "skill": 0.8502, "coverage": 1.0, "random": 0.5, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 21, "dataset": "Humicroedit", "requests": 2628, "answered": 2628, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 2628, "metric": "accuracy", "reference_same_cases": null, "index_raw": 0.6416, "index_skill": 0.2831, "coverage": 1.0, "chance": 0.5, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2628

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · Humicroedit · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.6416, "skill": 0.2831, "coverage": 1.0, "random": 0.5, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2628

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · Humicroedit · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.6416, "skill": 0.2831, "coverage": 1.0, "random": 0.5, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2628

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 22, "dataset": "POP909-CL", "requests": 2000, "answered": 2000, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 2000, "metric": "accuracy", "reference_same_cases": null, "index_raw": 0.2671, "index_skill": 0.2613, "coverage": 1.0, "chance": 0.0078, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · POP909-CL · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.2671, "skill": 0.2613, "coverage": 1.0, "random": 0.0078, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · POP909-CL · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.2671, "skill": 0.2613, "coverage": 1.0, "random": 0.0078, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 23, "dataset": "cfcolor", "requests": 5000, "answered": 5000, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 5000, "metric": "accuracy", "reference_same_cases": null, "index_raw": 0.6527, "index_skill": 0.3053, "coverage": 1.0, "chance": 0.5, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · cfcolor · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.6527, "skill": 0.3053, "coverage": 1.0, "random": 0.5, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · cfcolor · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.6527, "skill": 0.3053, "coverage": 1.0, "random": 0.5, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 24, "dataset": "MMLU", "requests": 14033, "answered": 14033, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 14033, "metric": "accuracy", "reference_same_cases": null, "index_raw": 0.8686, "index_skill": 0.8248, "coverage": 1.0, "chance": 0.25, "in_index": false}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 14033

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · MMLU · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.8686, "skill": 0.8248, "coverage": 1.0, "random": 0.25, "rule": "shown, not counted", "in_index": false}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 14033

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · MMLU · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.8686, "skill": 0.8248, "coverage": 1.0, "random": 0.25, "rule": "shown, not counted", "in_index": false}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 14033

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 25, "dataset": "GPQA Diamond", "requests": 196, "answered": 196, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 196, "metric": "accuracy", "reference_same_cases": null, "index_raw": 0.4949, "index_skill": 0.3265, "coverage": 1.0, "chance": 0.25, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 196

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · GPQA Diamond · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.4949, "skill": 0.3265, "coverage": 1.0, "random": 0.25, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 196

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · GPQA Diamond · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.4949, "skill": 0.3265, "coverage": 1.0, "random": 0.25, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 196

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 26, "dataset": "ARC-Easy", "requests": 2376, "answered": 2376, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 2376, "metric": "accuracy", "reference_same_cases": null, "index_raw": 0.9916, "index_skill": 0.9888, "coverage": 1.0, "chance": 0.2502, "in_index": false}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2376

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · ARC-Easy · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.9916, "skill": 0.9888, "coverage": 1.0, "random": 0.2502, "rule": "shown, not counted", "in_index": false}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2376

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · ARC-Easy · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.9916, "skill": 0.9888, "coverage": 1.0, "random": 0.2502, "rule": "shown, not counted", "in_index": false}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2376

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 27, "dataset": "ARC-Challenge", "requests": 1172, "answered": 1172, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 1172, "metric": "accuracy", "reference_same_cases": null, "index_raw": 0.9744, "index_skill": 0.9659, "coverage": 1.0, "chance": 0.2502, "in_index": false}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1172

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · ARC-Challenge · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.9744, "skill": 0.9659, "coverage": 1.0, "random": 0.2502, "rule": "shown, not counted", "in_index": false}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1172

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · ARC-Challenge · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.9744, "skill": 0.9659, "coverage": 1.0, "random": 0.2502, "rule": "shown, not counted", "in_index": false}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1172

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 28, "dataset": "WinoGrande", "requests": 1267, "answered": 1267, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 1267, "metric": "accuracy", "reference_same_cases": null, "index_raw": 0.9234, "index_skill": 0.8468, "coverage": 1.0, "chance": 0.5, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1267

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · WinoGrande · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.9234, "skill": 0.8468, "coverage": 1.0, "random": 0.5, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1267

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · WinoGrande · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.9234, "skill": 0.8468, "coverage": 1.0, "random": 0.5, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1267

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 29, "dataset": "HellaSwag", "requests": 10042, "answered": 10042, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 10042, "metric": "accuracy", "reference_same_cases": null, "index_raw": 0.9618, "index_skill": 0.9491, "coverage": 1.0, "chance": 0.25, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 10042

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · HellaSwag · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.9618, "skill": 0.9491, "coverage": 1.0, "random": 0.25, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 10042

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · HellaSwag · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.9618, "skill": 0.9491, "coverage": 1.0, "random": 0.25, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 10042

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 31, "dataset": "ChessBench", "requests": 5000, "answered": 5000, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 5000, "metric": "accuracy", "reference_same_cases": null, "index_raw": 0.236, "index_skill": 0.1679, "coverage": 1.0, "chance": 0.0819, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · ChessBench · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.236, "skill": 0.1679, "coverage": 1.0, "random": 0.0819, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · ChessBench · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.236, "skill": 0.1679, "coverage": 1.0, "random": 0.0819, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 32, "dataset": "MuSR", "requests": 752, "answered": 752, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 752, "metric": "accuracy", "reference_same_cases": null, "index_raw": 0.6503, "index_skill": 0.444, "coverage": 1.0, "chance": 0.371, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 752

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · MuSR · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.6503, "skill": 0.444, "coverage": 1.0, "random": 0.371, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 752

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · MuSR · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.6503, "skill": 0.444, "coverage": 1.0, "random": 0.371, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 752

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. case exact accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 33, "dataset": "SATA-Bench", "requests": 1650, "answered": 1650, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 1650, "metric": "case exact accuracy", "reference_same_cases": null, "index_raw": 0.1145, "index_skill": 0.1027, "coverage": 1.0, "chance": 0.0131, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1650

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · SATA-Bench · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.1145, "skill": 0.1027, "coverage": 1.0, "random": 0.0131, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1650

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · SATA-Bench · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.1145, "skill": 0.1027, "coverage": 1.0, "random": 0.0131, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1650

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 34, "dataset": "SimpleBench", "requests": 10, "answered": 10, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 10, "metric": "accuracy", "reference_same_cases": null, "index_raw": 0.2, "index_skill": 0.04, "coverage": 1.0, "chance": 0.1667, "in_index": false}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 10

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · SimpleBench · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.2, "skill": 0.04, "coverage": 1.0, "random": 0.1667, "rule": "shown, not counted", "in_index": false}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 10

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · SimpleBench · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.2, "skill": 0.04, "coverage": 1.0, "random": 0.1667, "rule": "shown, not counted", "in_index": false}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 10

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. nDCG@10

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 36, "dataset": "BRIGHT", "requests": 220, "answered": 220, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 220, "metric": "nDCG@10", "reference_same_cases": null, "index_raw": 0.4617, "index_skill": 0.391, "coverage": 1.0, "chance": 0.116, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 220

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · BRIGHT · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.4617, "skill": 0.391, "coverage": 1.0, "random": 0.116, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 220

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · BRIGHT · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.4617, "skill": 0.391, "coverage": 1.0, "random": 0.116, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 220

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. macro-F1

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 37, "dataset": "Amazon ESCI", "requests": 5000, "answered": 5000, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 5000, "metric": "macro-F1", "reference_same_cases": null, "index_raw": 0.5731, "index_skill": 0.4645, "coverage": 1.0, "chance": 0.2027, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · Amazon ESCI · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.5731, "skill": 0.4645, "coverage": 1.0, "random": 0.2027, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · Amazon ESCI · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.5731, "skill": 0.4645, "coverage": 1.0, "random": 0.2027, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. per-review F1

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 38, "dataset": "ACOS", "requests": 1565, "answered": 1565, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 1565, "metric": "per-review F1", "reference_same_cases": null, "index_raw": 0.2641, "index_skill": 0.2406, "coverage": 1.0, "chance": 0.031, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1565

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · ACOS · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.2641, "skill": 0.2406, "coverage": 1.0, "random": 0.031, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1565

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · ACOS · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.2641, "skill": 0.2406, "coverage": 1.0, "random": 0.031, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1565

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. macro-F1

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 39, "dataset": "FinEntity", "requests": 979, "answered": 979, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 979, "metric": "macro-F1", "reference_same_cases": null, "index_raw": 0.9292, "index_skill": 0.8959, "coverage": 1.0, "chance": 0.3201, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 979

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · FinEntity · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.9292, "skill": 0.8959, "coverage": 1.0, "random": 0.3201, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 979

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · FinEntity · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.9292, "skill": 0.8959, "coverage": 1.0, "random": 0.3201, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 979

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. Sarcasm F1 · track A, English

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 40, "dataset": "iSarcasmEval", "requests": 4600, "answered": 4600, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 4600, "metric": "Sarcasm F1 · track A, English", "reference_same_cases": null, "index_raw": 0.6189, "index_skill": 0.5097, "coverage": 1.0, "chance": 0.2227, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 4600

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · iSarcasmEval · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.6189, "skill": 0.5097, "coverage": 1.0, "random": 0.2227, "rule": "track", "in_index": true, "tracks": [{"track": "A · Arabic", "score": 0.6054, "headline": false}, {"track": "A · English", "score": 0.6189, "headline": true}, {"track": "C · Arabic pairs", "score": 0.975, "headline": false}, {"track": "C · English pairs", "score": 0.99, "headline": false}]}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 4600

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · iSarcasmEval · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.6189, "skill": 0.5097, "coverage": 1.0, "random": 0.2227, "rule": "track", "in_index": true, "tracks": [{"track": "A · Arabic", "score": 0.6054, "headline": false}, {"track": "A · English", "score": 0.6189, "headline": true}, {"track": "C · Arabic pairs", "score": 0.975, "headline": false}, {"track": "C · English pairs", "score": 0.99, "headline": false}]}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 4600

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · iSarcasmEval · A · Arabic

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Separate subtask/track; source track support {"track": "A · Arabic", "score": 0.6054, "headline": false}.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: Not published

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · iSarcasmEval · C · Arabic pairs

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Separate subtask/track; source track support {"track": "C · Arabic pairs", "score": 0.975, "headline": false}.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: Not published

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · iSarcasmEval · C · English pairs

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Separate subtask/track; source track support {"track": "C · English pairs", "score": 0.99, "headline": false}.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: Not published

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. macro-F1

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 41, "dataset": "VAST", "requests": 3006, "answered": 3006, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 3006, "metric": "macro-F1", "reference_same_cases": null, "index_raw": 0.7315, "index_skill": 0.5973, "coverage": 1.0, "chance": 0.3333, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 3006

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · VAST · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.7315, "skill": 0.5973, "coverage": 1.0, "random": 0.3333, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 3006

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · VAST · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.7315, "skill": 0.5973, "coverage": 1.0, "random": 0.3333, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 3006

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. macro-F1

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 42, "dataset": "NLI4CT", "requests": 5500, "answered": 5500, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 5500, "metric": "macro-F1", "reference_same_cases": null, "index_raw": 0.8441, "index_skill": 0.6967, "coverage": 1.0, "chance": 0.486, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5500

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · NLI4CT · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.8441, "skill": 0.6967, "coverage": 1.0, "random": 0.486, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5500

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · NLI4CT · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.8441, "skill": 0.6967, "coverage": 1.0, "random": 0.486, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5500

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 43, "dataset": "CRUXEval", "requests": 570, "answered": 570, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 570, "metric": "accuracy", "reference_same_cases": null, "index_raw": 0.8386, "index_skill": 0.7439, "coverage": 1.0, "chance": 0.3697, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 570

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · CRUXEval · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.8386, "skill": 0.7439, "coverage": 1.0, "random": 0.3697, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 570

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · CRUXEval · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.8386, "skill": 0.7439, "coverage": 1.0, "random": 0.3697, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 570

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 44, "dataset": "CLadder", "requests": 5000, "answered": 5000, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 5000, "metric": "accuracy", "reference_same_cases": null, "index_raw": 0.8016, "index_skill": 0.6032, "coverage": 1.0, "chance": 0.5, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · CLadder · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.8016, "skill": 0.6032, "coverage": 1.0, "random": 0.5, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · CLadder · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.8016, "skill": 0.6032, "coverage": 1.0, "random": 0.5, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 45, "dataset": "HLE", "requests": 501, "answered": 501, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 501, "metric": "accuracy", "reference_same_cases": null, "index_raw": 0.1018, "index_skill": 0.0, "coverage": 1.0, "chance": 0.1641, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 501

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · HLE · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.1018, "skill": 0.0, "coverage": 1.0, "random": 0.1641, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 501

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · HLE · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.1018, "skill": 0.0, "coverage": 1.0, "random": 0.1641, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 501

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 50, "dataset": "Habermas Machine", "requests": 1676, "answered": 1676, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 1676, "metric": "accuracy", "reference_same_cases": null, "index_raw": 0.4433, "index_skill": 0.1921, "coverage": 1.0, "chance": 0.311, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1676

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · Habermas Machine · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.4433, "skill": 0.1921, "coverage": 1.0, "random": 0.311, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1676

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · Habermas Machine · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.4433, "skill": 0.1921, "coverage": 1.0, "random": 0.311, "rule": "track", "in_index": true, "tracks": []}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1676

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 30, "dataset": "GSM8K", "requests": 2638, "answered": 2638, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "scored_requests": 2638, "metric": "accuracy", "reference_same_cases": null, "index_raw": 0.8624, "index_skill": 0.8346, "coverage": 1.0, "chance": 0.25, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2638

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · GSM8K · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.8624, "skill": 0.8346, "coverage": 1.0, "random": 0.25, "rule": "track", "in_index": true, "tracks": [{"track": "GSM8K-4choice", "score": 0.8878, "headline": false}, {"track": "GSM8K-10choice", "score": 0.837, "headline": false}]}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2638

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · GSM8K · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.8624, "skill": 0.8346, "coverage": 1.0, "random": 0.25, "rule": "track", "in_index": true, "tracks": [{"track": "GSM8K-4choice", "score": 0.8878, "headline": false}, {"track": "GSM8K-10choice", "score": 0.837, "headline": false}]}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2638

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · GSM8K · GSM8K-4choice

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Separate subtask/track; source track support {"track": "GSM8K-4choice", "score": 0.8878, "headline": false}.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: Not published

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · GSM8K · GSM8K-10choice

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Separate subtask/track; source track support {"track": "GSM8K-10choice", "score": 0.837, "headline": false}.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: Not published

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 56, "dataset": "PhishNChips phishing decisions", "requests": 2000, "answered": 2000, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "metric": "accuracy", "scored_requests": 2000, "index_raw": 0.9245, "index_skill": 0.849, "coverage": 1.0, "chance": 0.5, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · PhishNChips phishing decisions · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.9245, "skill": 0.849, "coverage": 1.0, "random": 0.5, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · PhishNChips phishing decisions · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.9245, "skill": 0.849, "coverage": 1.0, "random": 0.5, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 57, "dataset": "MMLU-Pro", "requests": 12032, "answered": 12032, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "metric": "accuracy", "scored_requests": 12032, "index_raw": 0.7197, "index_skill": 0.6847, "coverage": 1.0, "chance": 0.1109, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 12032

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · MMLU-Pro · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.7197, "skill": 0.6847, "coverage": 1.0, "random": 0.1109, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 12032

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · MMLU-Pro · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.7197, "skill": 0.6847, "coverage": 1.0, "random": 0.1109, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 12032

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 58, "dataset": "BBH fixed-option tasks", "requests": 5507, "answered": 5507, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "metric": "accuracy", "scored_requests": 5507, "index_raw": 0.8064, "index_skill": 0.7194, "coverage": 1.0, "chance": 0.3101, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5507

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · BBH fixed-option tasks · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.8064, "skill": 0.7194, "coverage": 1.0, "random": 0.3101, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5507

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · BBH fixed-option tasks · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.8064, "skill": 0.7194, "coverage": 1.0, "random": 0.3101, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5507

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. F1 on hallucinated class

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 59, "dataset": "RAGTruth response-level hallucination", "requests": 2700, "answered": 2700, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "metric": "F1 on hallucinated class", "scored_requests": 2700, "index_raw": 0.7944, "index_skill": 0.5737, "coverage": 1.0, "chance": 0.5177, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2700

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · RAGTruth response-level hallucination · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.7944, "skill": 0.5737, "coverage": 1.0, "random": 0.5177, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2700

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · RAGTruth response-level hallucination · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.7944, "skill": 0.5737, "coverage": 1.0, "random": 0.5177, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2700

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 61, "dataset": "HoVer claim verification", "requests": 4000, "answered": 4000, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "metric": "accuracy", "scored_requests": 4000, "index_raw": 0.83, "index_skill": 0.66, "coverage": 1.0, "chance": 0.5, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 4000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · HoVer claim verification · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.83, "skill": 0.66, "coverage": 1.0, "random": 0.5, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 4000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · HoVer claim verification · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.83, "skill": 0.66, "coverage": 1.0, "random": 0.5, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 4000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 62, "dataset": "When2Call MCQ", "requests": 3652, "answered": 3652, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "metric": "accuracy", "scored_requests": 3652, "index_raw": 0.8436, "index_skill": 0.7915, "coverage": 1.0, "chance": 0.25, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 3652

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · When2Call MCQ · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.8436, "skill": 0.7915, "coverage": 1.0, "random": 0.25, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 3652

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · When2Call MCQ · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.8436, "skill": 0.7915, "coverage": 1.0, "random": 0.25, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 3652

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Native source metric on this vendor submission population. accuracy

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Support/errors/status: {"catalog_id": 64, "dataset": "New Yorker caption matching", "requests": 528, "answered": 528, "unsupported": 0, "errors": 0, "abstained": 0, "pending": 0, "metric": "accuracy", "scored_requests": 528, "index_raw": 0.7746, "index_skill": 0.7182, "coverage": 1.0, "chance": 0.2, "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 528

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · New Yorker caption matching · Raw index component

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.7746, "skill": 0.7182, "coverage": 1.0, "random": 0.2, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 528

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Sieve vendor DI 0.3 · New Yorker caption matching · Chance-corrected public skill

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Source index convention may differ from native score (e.g. cluster-macro). Exact normalization: {"raw": 0.7746, "skill": 0.7182, "coverage": 1.0, "random": 0.2, "rule": "chance", "in_index": true}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 528

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary field_accuracy on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.field_accuracy; whole detail record: {"field_accuracy": 0.9890939597315436, "case_exact_accuracy": 0.9775678866587958, "custom_metrics": {}, "positive_micro_f1": 0.9846335697399528}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1694

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary positive_micro_f1 on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.positive_micro_f1; whole detail record: {"field_accuracy": 0.9890939597315436, "case_exact_accuracy": 0.9775678866587958, "custom_metrics": {}, "positive_micro_f1": 0.9846335697399528}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1694

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary custom_metrics / mrr on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.custom_metrics.mrr; whole detail record: {"field_accuracy": null, "case_exact_accuracy": null, "custom_metrics": {"ndcg_at_10": 0.6751653602081698, "mrr": 0.7126938666191451, "recall_at_10": 0.7972766770941954, "candidate_recall": 0.8401355578727842, "scorable_candidate_recall": 0.8401355578727842, "candidates_scored": 31.9985401459854, "candidates_retrieved": 32, "bm25_ndcg_at_10": 0.5430661225020075}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 685

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary custom_metrics / recall_at_10 on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.custom_metrics.recall_at_10; whole detail record: {"field_accuracy": null, "case_exact_accuracy": null, "custom_metrics": {"ndcg_at_10": 0.6751653602081698, "mrr": 0.7126938666191451, "recall_at_10": 0.7972766770941954, "candidate_recall": 0.8401355578727842, "scorable_candidate_recall": 0.8401355578727842, "candidates_scored": 31.9985401459854, "candidates_retrieved": 32, "bm25_ndcg_at_10": 0.5430661225020075}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 685

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary custom_metrics / candidate_recall on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.custom_metrics.candidate_recall; whole detail record: {"field_accuracy": null, "case_exact_accuracy": null, "custom_metrics": {"ndcg_at_10": 0.6751653602081698, "mrr": 0.7126938666191451, "recall_at_10": 0.7972766770941954, "candidate_recall": 0.8401355578727842, "scorable_candidate_recall": 0.8401355578727842, "candidates_scored": 31.9985401459854, "candidates_retrieved": 32, "bm25_ndcg_at_10": 0.5430661225020075}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 685

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary custom_metrics / scorable_candidate_recall on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.custom_metrics.scorable_candidate_recall; whole detail record: {"field_accuracy": null, "case_exact_accuracy": null, "custom_metrics": {"ndcg_at_10": 0.6751653602081698, "mrr": 0.7126938666191451, "recall_at_10": 0.7972766770941954, "candidate_recall": 0.8401355578727842, "scorable_candidate_recall": 0.8401355578727842, "candidates_scored": 31.9985401459854, "candidates_retrieved": 32, "bm25_ndcg_at_10": 0.5430661225020075}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 685

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary field_accuracy on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.field_accuracy; whole detail record: {"field_accuracy": 0.8613636363636363, "case_exact_accuracy": 0.8613636363636363, "macro_f1": 0.85645362658449, "custom_metrics": {}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 3080

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary case_exact_accuracy on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.case_exact_accuracy; whole detail record: {"field_accuracy": 0.8613636363636363, "case_exact_accuracy": 0.8613636363636363, "macro_f1": 0.85645362658449, "custom_metrics": {}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 3080

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary field_accuracy on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.field_accuracy; whole detail record: {"field_accuracy": 0.936, "case_exact_accuracy": 0.936, "macro_f1": 0.9354863295310092, "custom_metrics": {}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5500

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary case_exact_accuracy on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.case_exact_accuracy; whole detail record: {"field_accuracy": 0.936, "case_exact_accuracy": 0.936, "macro_f1": 0.9354863295310092, "custom_metrics": {}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5500

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary custom_metrics / quality_quality on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.custom_metrics.quality_quality; whole detail record: {"field_accuracy": null, "case_exact_accuracy": null, "custom_metrics": {"quality_quality": 0.799775, "quality_cost_usd": 0.00539312831, "quality_utility": 0.799775, "quality_oracle_optimal": 0.8354, "quality_utility_regret": 0.119285, "cost_aware_quality": 0.73941, "cost_aware_cost_usd": 0.00144725863, "cost_aware_utility": 0.7249374137, "cost_aware_oracle_optimal": 0.0995, "cost_aware_utility_regret": 0.190591928928}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 10000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary custom_metrics / quality_cost_usd on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.custom_metrics.quality_cost_usd; whole detail record: {"field_accuracy": null, "case_exact_accuracy": null, "custom_metrics": {"quality_quality": 0.799775, "quality_cost_usd": 0.00539312831, "quality_utility": 0.799775, "quality_oracle_optimal": 0.8354, "quality_utility_regret": 0.119285, "cost_aware_quality": 0.73941, "cost_aware_cost_usd": 0.00144725863, "cost_aware_utility": 0.7249374137, "cost_aware_oracle_optimal": 0.0995, "cost_aware_utility_regret": 0.190591928928}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 10000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary custom_metrics / quality_utility on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.custom_metrics.quality_utility; whole detail record: {"field_accuracy": null, "case_exact_accuracy": null, "custom_metrics": {"quality_quality": 0.799775, "quality_cost_usd": 0.00539312831, "quality_utility": 0.799775, "quality_oracle_optimal": 0.8354, "quality_utility_regret": 0.119285, "cost_aware_quality": 0.73941, "cost_aware_cost_usd": 0.00144725863, "cost_aware_utility": 0.7249374137, "cost_aware_oracle_optimal": 0.0995, "cost_aware_utility_regret": 0.190591928928}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 10000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary custom_metrics / quality_oracle_optimal on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.custom_metrics.quality_oracle_optimal; whole detail record: {"field_accuracy": null, "case_exact_accuracy": null, "custom_metrics": {"quality_quality": 0.799775, "quality_cost_usd": 0.00539312831, "quality_utility": 0.799775, "quality_oracle_optimal": 0.8354, "quality_utility_regret": 0.119285, "cost_aware_quality": 0.73941, "cost_aware_cost_usd": 0.00144725863, "cost_aware_utility": 0.7249374137, "cost_aware_oracle_optimal": 0.0995, "cost_aware_utility_regret": 0.190591928928}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 10000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary custom_metrics / quality_utility_regret on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.custom_metrics.quality_utility_regret; whole detail record: {"field_accuracy": null, "case_exact_accuracy": null, "custom_metrics": {"quality_quality": 0.799775, "quality_cost_usd": 0.00539312831, "quality_utility": 0.799775, "quality_oracle_optimal": 0.8354, "quality_utility_regret": 0.119285, "cost_aware_quality": 0.73941, "cost_aware_cost_usd": 0.00144725863, "cost_aware_utility": 0.7249374137, "cost_aware_oracle_optimal": 0.0995, "cost_aware_utility_regret": 0.190591928928}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 10000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary custom_metrics / cost_aware_quality on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.custom_metrics.cost_aware_quality; whole detail record: {"field_accuracy": null, "case_exact_accuracy": null, "custom_metrics": {"quality_quality": 0.799775, "quality_cost_usd": 0.00539312831, "quality_utility": 0.799775, "quality_oracle_optimal": 0.8354, "quality_utility_regret": 0.119285, "cost_aware_quality": 0.73941, "cost_aware_cost_usd": 0.00144725863, "cost_aware_utility": 0.7249374137, "cost_aware_oracle_optimal": 0.0995, "cost_aware_utility_regret": 0.190591928928}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 10000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary custom_metrics / cost_aware_cost_usd on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.custom_metrics.cost_aware_cost_usd; whole detail record: {"field_accuracy": null, "case_exact_accuracy": null, "custom_metrics": {"quality_quality": 0.799775, "quality_cost_usd": 0.00539312831, "quality_utility": 0.799775, "quality_oracle_optimal": 0.8354, "quality_utility_regret": 0.119285, "cost_aware_quality": 0.73941, "cost_aware_cost_usd": 0.00144725863, "cost_aware_utility": 0.7249374137, "cost_aware_oracle_optimal": 0.0995, "cost_aware_utility_regret": 0.190591928928}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 10000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary custom_metrics / cost_aware_utility on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.custom_metrics.cost_aware_utility; whole detail record: {"field_accuracy": null, "case_exact_accuracy": null, "custom_metrics": {"quality_quality": 0.799775, "quality_cost_usd": 0.00539312831, "quality_utility": 0.799775, "quality_oracle_optimal": 0.8354, "quality_utility_regret": 0.119285, "cost_aware_quality": 0.73941, "cost_aware_cost_usd": 0.00144725863, "cost_aware_utility": 0.7249374137, "cost_aware_oracle_optimal": 0.0995, "cost_aware_utility_regret": 0.190591928928}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 10000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary custom_metrics / cost_aware_oracle_optimal on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.custom_metrics.cost_aware_oracle_optimal; whole detail record: {"field_accuracy": null, "case_exact_accuracy": null, "custom_metrics": {"quality_quality": 0.799775, "quality_cost_usd": 0.00539312831, "quality_utility": 0.799775, "quality_oracle_optimal": 0.8354, "quality_utility_regret": 0.119285, "cost_aware_quality": 0.73941, "cost_aware_cost_usd": 0.00144725863, "cost_aware_utility": 0.7249374137, "cost_aware_oracle_optimal": 0.0995, "cost_aware_utility_regret": 0.190591928928}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 10000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary custom_metrics / cost_aware_utility_regret on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.custom_metrics.cost_aware_utility_regret; whole detail record: {"field_accuracy": null, "case_exact_accuracy": null, "custom_metrics": {"quality_quality": 0.799775, "quality_cost_usd": 0.00539312831, "quality_utility": 0.799775, "quality_oracle_optimal": 0.8354, "quality_utility_regret": 0.119285, "cost_aware_quality": 0.73941, "cost_aware_cost_usd": 0.00144725863, "cost_aware_utility": 0.7249374137, "cost_aware_oracle_optimal": 0.0995, "cost_aware_utility_regret": 0.190591928928}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 10000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary field_accuracy on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.field_accuracy; whole detail record: {"field_accuracy": 0.969811320754717, "case_exact_accuracy": 0.625, "custom_metrics": {}, "subgroups": {"outcome": {"n": 88, "mean": 0.9204545454545454}, "post_0": {"n": 69, "mean": 1}, "post_1": {"n": 25, "mean": 0.88}, "resolution": {"n": 88, "mean": 0.8636363636363636}, "state_read": {"n": 88, "mean": 1}, "target_member": {"n": 1232, "mean": 0.9788961038961039}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 88

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / outcome / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.outcome.mean; whole detail record: {"field_accuracy": 0.969811320754717, "case_exact_accuracy": 0.625, "custom_metrics": {}, "subgroups": {"outcome": {"n": 88, "mean": 0.9204545454545454}, "post_0": {"n": 69, "mean": 1}, "post_1": {"n": 25, "mean": 0.88}, "resolution": {"n": 88, "mean": 0.8636363636363636}, "state_read": {"n": 88, "mean": 1}, "target_member": {"n": 1232, "mean": 0.9788961038961039}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 88

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / post_0 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.post_0.mean; whole detail record: {"field_accuracy": 0.969811320754717, "case_exact_accuracy": 0.625, "custom_metrics": {}, "subgroups": {"outcome": {"n": 88, "mean": 0.9204545454545454}, "post_0": {"n": 69, "mean": 1}, "post_1": {"n": 25, "mean": 0.88}, "resolution": {"n": 88, "mean": 0.8636363636363636}, "state_read": {"n": 88, "mean": 1}, "target_member": {"n": 1232, "mean": 0.9788961038961039}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 69

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / post_1 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.post_1.mean; whole detail record: {"field_accuracy": 0.969811320754717, "case_exact_accuracy": 0.625, "custom_metrics": {}, "subgroups": {"outcome": {"n": 88, "mean": 0.9204545454545454}, "post_0": {"n": 69, "mean": 1}, "post_1": {"n": 25, "mean": 0.88}, "resolution": {"n": 88, "mean": 0.8636363636363636}, "state_read": {"n": 88, "mean": 1}, "target_member": {"n": 1232, "mean": 0.9788961038961039}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 25

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / resolution / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.resolution.mean; whole detail record: {"field_accuracy": 0.969811320754717, "case_exact_accuracy": 0.625, "custom_metrics": {}, "subgroups": {"outcome": {"n": 88, "mean": 0.9204545454545454}, "post_0": {"n": 69, "mean": 1}, "post_1": {"n": 25, "mean": 0.88}, "resolution": {"n": 88, "mean": 0.8636363636363636}, "state_read": {"n": 88, "mean": 1}, "target_member": {"n": 1232, "mean": 0.9788961038961039}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 88

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / state_read / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.state_read.mean; whole detail record: {"field_accuracy": 0.969811320754717, "case_exact_accuracy": 0.625, "custom_metrics": {}, "subgroups": {"outcome": {"n": 88, "mean": 0.9204545454545454}, "post_0": {"n": 69, "mean": 1}, "post_1": {"n": 25, "mean": 0.88}, "resolution": {"n": 88, "mean": 0.8636363636363636}, "state_read": {"n": 88, "mean": 1}, "target_member": {"n": 1232, "mean": 0.9788961038961039}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 88

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / target_member / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.target_member.mean; whole detail record: {"field_accuracy": 0.969811320754717, "case_exact_accuracy": 0.625, "custom_metrics": {}, "subgroups": {"outcome": {"n": 88, "mean": 0.9204545454545454}, "post_0": {"n": 69, "mean": 1}, "post_1": {"n": 25, "mean": 0.88}, "resolution": {"n": 88, "mean": 0.8636363636363636}, "state_read": {"n": 88, "mean": 1}, "target_member": {"n": 1232, "mean": 0.9788961038961039}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1232

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary field_accuracy on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.field_accuracy; whole detail record: {"field_accuracy": 0.2908, "case_exact_accuracy": 0.2908, "macro_f1": 0.3219005868847953, "custom_metrics": {}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2500

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary case_exact_accuracy on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.case_exact_accuracy; whole detail record: {"field_accuracy": 0.2908, "case_exact_accuracy": 0.2908, "macro_f1": 0.3219005868847953, "custom_metrics": {}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2500

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary field_accuracy on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.field_accuracy; whole detail record: {"field_accuracy": 0.8546150167384027, "case_exact_accuracy": 0.04878048780487805, "macro_f1": 0.8125468081507672, "custom_metrics": {}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 123

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary case_exact_accuracy on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.case_exact_accuracy; whole detail record: {"field_accuracy": 0.8546150167384027, "case_exact_accuracy": 0.04878048780487805, "macro_f1": 0.8125468081507672, "custom_metrics": {}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 123

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary field_accuracy on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.field_accuracy; whole detail record: {"field_accuracy": 0.768125, "case_exact_accuracy": 0.768125, "macro_f1": 0.7686474147925114, "custom_metrics": {}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 3200

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary case_exact_accuracy on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.case_exact_accuracy; whole detail record: {"field_accuracy": 0.768125, "case_exact_accuracy": 0.768125, "macro_f1": 0.7686474147925114, "custom_metrics": {}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 3200

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary case_exact_accuracy on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.case_exact_accuracy; whole detail record: {"field_accuracy": 0.9228, "case_exact_accuracy": 0.6707768187422934, "custom_metrics": {}, "subgroups": {"beam search": {"n": 401, "mean": 0.830423940149626}, "corrupted rhyme": {"n": 376, "mean": 0.9840425531914894}, "delayed beam search": {"n": 137, "mean": 0.9927007299270073}, "delete any 1 nonrhyming word": {"n": 376, "mean": 0.8962765957446809}, "delete any 1 rhyming EOL word": {"n": 376, "mean": 0.9893617021276596}, "delete any 1 word": {"n": 376, "mean": 0.8962765957446809}, "delete any 2 nonrhyming words": {"n": 376, "mean": 0.9148936170212766}, "delete any 2 rhyming EOL words": {"n": 376, "mean": 0.9920212765957447}, "delete any 2 words": {"n": 376, "mean": 0.9468085106382979}, "delete any 3 nonrhyming words": {"n": 376, "mean": 0.9095744680851063}, "delete any 3 rhyming words": {"n": 376, "mean": 0.9840425531914894}, "delete any 3 words": {"n": 376, "mean": 0.949468085106383}, "shuffled rhyming lines": {"n": 336, "mean": 0.7529761904761905}, "shuffled rhyming words": {"n": 366, "mean": 0.912568306010929}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / beam search / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.beam search.mean; whole detail record: {"field_accuracy": 0.9228, "case_exact_accuracy": 0.6707768187422934, "custom_metrics": {}, "subgroups": {"beam search": {"n": 401, "mean": 0.830423940149626}, "corrupted rhyme": {"n": 376, "mean": 0.9840425531914894}, "delayed beam search": {"n": 137, "mean": 0.9927007299270073}, "delete any 1 nonrhyming word": {"n": 376, "mean": 0.8962765957446809}, "delete any 1 rhyming EOL word": {"n": 376, "mean": 0.9893617021276596}, "delete any 1 word": {"n": 376, "mean": 0.8962765957446809}, "delete any 2 nonrhyming words": {"n": 376, "mean": 0.9148936170212766}, "delete any 2 rhyming EOL words": {"n": 376, "mean": 0.9920212765957447}, "delete any 2 words": {"n": 376, "mean": 0.9468085106382979}, "delete any 3 nonrhyming words": {"n": 376, "mean": 0.9095744680851063}, "delete any 3 rhyming words": {"n": 376, "mean": 0.9840425531914894}, "delete any 3 words": {"n": 376, "mean": 0.949468085106383}, "shuffled rhyming lines": {"n": 336, "mean": 0.7529761904761905}, "shuffled rhyming words": {"n": 366, "mean": 0.912568306010929}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 401

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / corrupted rhyme / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.corrupted rhyme.mean; whole detail record: {"field_accuracy": 0.9228, "case_exact_accuracy": 0.6707768187422934, "custom_metrics": {}, "subgroups": {"beam search": {"n": 401, "mean": 0.830423940149626}, "corrupted rhyme": {"n": 376, "mean": 0.9840425531914894}, "delayed beam search": {"n": 137, "mean": 0.9927007299270073}, "delete any 1 nonrhyming word": {"n": 376, "mean": 0.8962765957446809}, "delete any 1 rhyming EOL word": {"n": 376, "mean": 0.9893617021276596}, "delete any 1 word": {"n": 376, "mean": 0.8962765957446809}, "delete any 2 nonrhyming words": {"n": 376, "mean": 0.9148936170212766}, "delete any 2 rhyming EOL words": {"n": 376, "mean": 0.9920212765957447}, "delete any 2 words": {"n": 376, "mean": 0.9468085106382979}, "delete any 3 nonrhyming words": {"n": 376, "mean": 0.9095744680851063}, "delete any 3 rhyming words": {"n": 376, "mean": 0.9840425531914894}, "delete any 3 words": {"n": 376, "mean": 0.949468085106383}, "shuffled rhyming lines": {"n": 336, "mean": 0.7529761904761905}, "shuffled rhyming words": {"n": 366, "mean": 0.912568306010929}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 376

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / delayed beam search / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.delayed beam search.mean; whole detail record: {"field_accuracy": 0.9228, "case_exact_accuracy": 0.6707768187422934, "custom_metrics": {}, "subgroups": {"beam search": {"n": 401, "mean": 0.830423940149626}, "corrupted rhyme": {"n": 376, "mean": 0.9840425531914894}, "delayed beam search": {"n": 137, "mean": 0.9927007299270073}, "delete any 1 nonrhyming word": {"n": 376, "mean": 0.8962765957446809}, "delete any 1 rhyming EOL word": {"n": 376, "mean": 0.9893617021276596}, "delete any 1 word": {"n": 376, "mean": 0.8962765957446809}, "delete any 2 nonrhyming words": {"n": 376, "mean": 0.9148936170212766}, "delete any 2 rhyming EOL words": {"n": 376, "mean": 0.9920212765957447}, "delete any 2 words": {"n": 376, "mean": 0.9468085106382979}, "delete any 3 nonrhyming words": {"n": 376, "mean": 0.9095744680851063}, "delete any 3 rhyming words": {"n": 376, "mean": 0.9840425531914894}, "delete any 3 words": {"n": 376, "mean": 0.949468085106383}, "shuffled rhyming lines": {"n": 336, "mean": 0.7529761904761905}, "shuffled rhyming words": {"n": 366, "mean": 0.912568306010929}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 137

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / delete any 1 nonrhyming word / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.delete any 1 nonrhyming word.mean; whole detail record: {"field_accuracy": 0.9228, "case_exact_accuracy": 0.6707768187422934, "custom_metrics": {}, "subgroups": {"beam search": {"n": 401, "mean": 0.830423940149626}, "corrupted rhyme": {"n": 376, "mean": 0.9840425531914894}, "delayed beam search": {"n": 137, "mean": 0.9927007299270073}, "delete any 1 nonrhyming word": {"n": 376, "mean": 0.8962765957446809}, "delete any 1 rhyming EOL word": {"n": 376, "mean": 0.9893617021276596}, "delete any 1 word": {"n": 376, "mean": 0.8962765957446809}, "delete any 2 nonrhyming words": {"n": 376, "mean": 0.9148936170212766}, "delete any 2 rhyming EOL words": {"n": 376, "mean": 0.9920212765957447}, "delete any 2 words": {"n": 376, "mean": 0.9468085106382979}, "delete any 3 nonrhyming words": {"n": 376, "mean": 0.9095744680851063}, "delete any 3 rhyming words": {"n": 376, "mean": 0.9840425531914894}, "delete any 3 words": {"n": 376, "mean": 0.949468085106383}, "shuffled rhyming lines": {"n": 336, "mean": 0.7529761904761905}, "shuffled rhyming words": {"n": 366, "mean": 0.912568306010929}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 376

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / delete any 1 rhyming EOL word / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.delete any 1 rhyming EOL word.mean; whole detail record: {"field_accuracy": 0.9228, "case_exact_accuracy": 0.6707768187422934, "custom_metrics": {}, "subgroups": {"beam search": {"n": 401, "mean": 0.830423940149626}, "corrupted rhyme": {"n": 376, "mean": 0.9840425531914894}, "delayed beam search": {"n": 137, "mean": 0.9927007299270073}, "delete any 1 nonrhyming word": {"n": 376, "mean": 0.8962765957446809}, "delete any 1 rhyming EOL word": {"n": 376, "mean": 0.9893617021276596}, "delete any 1 word": {"n": 376, "mean": 0.8962765957446809}, "delete any 2 nonrhyming words": {"n": 376, "mean": 0.9148936170212766}, "delete any 2 rhyming EOL words": {"n": 376, "mean": 0.9920212765957447}, "delete any 2 words": {"n": 376, "mean": 0.9468085106382979}, "delete any 3 nonrhyming words": {"n": 376, "mean": 0.9095744680851063}, "delete any 3 rhyming words": {"n": 376, "mean": 0.9840425531914894}, "delete any 3 words": {"n": 376, "mean": 0.949468085106383}, "shuffled rhyming lines": {"n": 336, "mean": 0.7529761904761905}, "shuffled rhyming words": {"n": 366, "mean": 0.912568306010929}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 376

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / delete any 1 word / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.delete any 1 word.mean; whole detail record: {"field_accuracy": 0.9228, "case_exact_accuracy": 0.6707768187422934, "custom_metrics": {}, "subgroups": {"beam search": {"n": 401, "mean": 0.830423940149626}, "corrupted rhyme": {"n": 376, "mean": 0.9840425531914894}, "delayed beam search": {"n": 137, "mean": 0.9927007299270073}, "delete any 1 nonrhyming word": {"n": 376, "mean": 0.8962765957446809}, "delete any 1 rhyming EOL word": {"n": 376, "mean": 0.9893617021276596}, "delete any 1 word": {"n": 376, "mean": 0.8962765957446809}, "delete any 2 nonrhyming words": {"n": 376, "mean": 0.9148936170212766}, "delete any 2 rhyming EOL words": {"n": 376, "mean": 0.9920212765957447}, "delete any 2 words": {"n": 376, "mean": 0.9468085106382979}, "delete any 3 nonrhyming words": {"n": 376, "mean": 0.9095744680851063}, "delete any 3 rhyming words": {"n": 376, "mean": 0.9840425531914894}, "delete any 3 words": {"n": 376, "mean": 0.949468085106383}, "shuffled rhyming lines": {"n": 336, "mean": 0.7529761904761905}, "shuffled rhyming words": {"n": 366, "mean": 0.912568306010929}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 376

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / delete any 2 nonrhyming words / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.delete any 2 nonrhyming words.mean; whole detail record: {"field_accuracy": 0.9228, "case_exact_accuracy": 0.6707768187422934, "custom_metrics": {}, "subgroups": {"beam search": {"n": 401, "mean": 0.830423940149626}, "corrupted rhyme": {"n": 376, "mean": 0.9840425531914894}, "delayed beam search": {"n": 137, "mean": 0.9927007299270073}, "delete any 1 nonrhyming word": {"n": 376, "mean": 0.8962765957446809}, "delete any 1 rhyming EOL word": {"n": 376, "mean": 0.9893617021276596}, "delete any 1 word": {"n": 376, "mean": 0.8962765957446809}, "delete any 2 nonrhyming words": {"n": 376, "mean": 0.9148936170212766}, "delete any 2 rhyming EOL words": {"n": 376, "mean": 0.9920212765957447}, "delete any 2 words": {"n": 376, "mean": 0.9468085106382979}, "delete any 3 nonrhyming words": {"n": 376, "mean": 0.9095744680851063}, "delete any 3 rhyming words": {"n": 376, "mean": 0.9840425531914894}, "delete any 3 words": {"n": 376, "mean": 0.949468085106383}, "shuffled rhyming lines": {"n": 336, "mean": 0.7529761904761905}, "shuffled rhyming words": {"n": 366, "mean": 0.912568306010929}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 376

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / delete any 2 rhyming EOL words / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.delete any 2 rhyming EOL words.mean; whole detail record: {"field_accuracy": 0.9228, "case_exact_accuracy": 0.6707768187422934, "custom_metrics": {}, "subgroups": {"beam search": {"n": 401, "mean": 0.830423940149626}, "corrupted rhyme": {"n": 376, "mean": 0.9840425531914894}, "delayed beam search": {"n": 137, "mean": 0.9927007299270073}, "delete any 1 nonrhyming word": {"n": 376, "mean": 0.8962765957446809}, "delete any 1 rhyming EOL word": {"n": 376, "mean": 0.9893617021276596}, "delete any 1 word": {"n": 376, "mean": 0.8962765957446809}, "delete any 2 nonrhyming words": {"n": 376, "mean": 0.9148936170212766}, "delete any 2 rhyming EOL words": {"n": 376, "mean": 0.9920212765957447}, "delete any 2 words": {"n": 376, "mean": 0.9468085106382979}, "delete any 3 nonrhyming words": {"n": 376, "mean": 0.9095744680851063}, "delete any 3 rhyming words": {"n": 376, "mean": 0.9840425531914894}, "delete any 3 words": {"n": 376, "mean": 0.949468085106383}, "shuffled rhyming lines": {"n": 336, "mean": 0.7529761904761905}, "shuffled rhyming words": {"n": 366, "mean": 0.912568306010929}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 376

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / delete any 2 words / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.delete any 2 words.mean; whole detail record: {"field_accuracy": 0.9228, "case_exact_accuracy": 0.6707768187422934, "custom_metrics": {}, "subgroups": {"beam search": {"n": 401, "mean": 0.830423940149626}, "corrupted rhyme": {"n": 376, "mean": 0.9840425531914894}, "delayed beam search": {"n": 137, "mean": 0.9927007299270073}, "delete any 1 nonrhyming word": {"n": 376, "mean": 0.8962765957446809}, "delete any 1 rhyming EOL word": {"n": 376, "mean": 0.9893617021276596}, "delete any 1 word": {"n": 376, "mean": 0.8962765957446809}, "delete any 2 nonrhyming words": {"n": 376, "mean": 0.9148936170212766}, "delete any 2 rhyming EOL words": {"n": 376, "mean": 0.9920212765957447}, "delete any 2 words": {"n": 376, "mean": 0.9468085106382979}, "delete any 3 nonrhyming words": {"n": 376, "mean": 0.9095744680851063}, "delete any 3 rhyming words": {"n": 376, "mean": 0.9840425531914894}, "delete any 3 words": {"n": 376, "mean": 0.949468085106383}, "shuffled rhyming lines": {"n": 336, "mean": 0.7529761904761905}, "shuffled rhyming words": {"n": 366, "mean": 0.912568306010929}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 376

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / delete any 3 nonrhyming words / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.delete any 3 nonrhyming words.mean; whole detail record: {"field_accuracy": 0.9228, "case_exact_accuracy": 0.6707768187422934, "custom_metrics": {}, "subgroups": {"beam search": {"n": 401, "mean": 0.830423940149626}, "corrupted rhyme": {"n": 376, "mean": 0.9840425531914894}, "delayed beam search": {"n": 137, "mean": 0.9927007299270073}, "delete any 1 nonrhyming word": {"n": 376, "mean": 0.8962765957446809}, "delete any 1 rhyming EOL word": {"n": 376, "mean": 0.9893617021276596}, "delete any 1 word": {"n": 376, "mean": 0.8962765957446809}, "delete any 2 nonrhyming words": {"n": 376, "mean": 0.9148936170212766}, "delete any 2 rhyming EOL words": {"n": 376, "mean": 0.9920212765957447}, "delete any 2 words": {"n": 376, "mean": 0.9468085106382979}, "delete any 3 nonrhyming words": {"n": 376, "mean": 0.9095744680851063}, "delete any 3 rhyming words": {"n": 376, "mean": 0.9840425531914894}, "delete any 3 words": {"n": 376, "mean": 0.949468085106383}, "shuffled rhyming lines": {"n": 336, "mean": 0.7529761904761905}, "shuffled rhyming words": {"n": 366, "mean": 0.912568306010929}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 376

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / delete any 3 rhyming words / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.delete any 3 rhyming words.mean; whole detail record: {"field_accuracy": 0.9228, "case_exact_accuracy": 0.6707768187422934, "custom_metrics": {}, "subgroups": {"beam search": {"n": 401, "mean": 0.830423940149626}, "corrupted rhyme": {"n": 376, "mean": 0.9840425531914894}, "delayed beam search": {"n": 137, "mean": 0.9927007299270073}, "delete any 1 nonrhyming word": {"n": 376, "mean": 0.8962765957446809}, "delete any 1 rhyming EOL word": {"n": 376, "mean": 0.9893617021276596}, "delete any 1 word": {"n": 376, "mean": 0.8962765957446809}, "delete any 2 nonrhyming words": {"n": 376, "mean": 0.9148936170212766}, "delete any 2 rhyming EOL words": {"n": 376, "mean": 0.9920212765957447}, "delete any 2 words": {"n": 376, "mean": 0.9468085106382979}, "delete any 3 nonrhyming words": {"n": 376, "mean": 0.9095744680851063}, "delete any 3 rhyming words": {"n": 376, "mean": 0.9840425531914894}, "delete any 3 words": {"n": 376, "mean": 0.949468085106383}, "shuffled rhyming lines": {"n": 336, "mean": 0.7529761904761905}, "shuffled rhyming words": {"n": 366, "mean": 0.912568306010929}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 376

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / delete any 3 words / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.delete any 3 words.mean; whole detail record: {"field_accuracy": 0.9228, "case_exact_accuracy": 0.6707768187422934, "custom_metrics": {}, "subgroups": {"beam search": {"n": 401, "mean": 0.830423940149626}, "corrupted rhyme": {"n": 376, "mean": 0.9840425531914894}, "delayed beam search": {"n": 137, "mean": 0.9927007299270073}, "delete any 1 nonrhyming word": {"n": 376, "mean": 0.8962765957446809}, "delete any 1 rhyming EOL word": {"n": 376, "mean": 0.9893617021276596}, "delete any 1 word": {"n": 376, "mean": 0.8962765957446809}, "delete any 2 nonrhyming words": {"n": 376, "mean": 0.9148936170212766}, "delete any 2 rhyming EOL words": {"n": 376, "mean": 0.9920212765957447}, "delete any 2 words": {"n": 376, "mean": 0.9468085106382979}, "delete any 3 nonrhyming words": {"n": 376, "mean": 0.9095744680851063}, "delete any 3 rhyming words": {"n": 376, "mean": 0.9840425531914894}, "delete any 3 words": {"n": 376, "mean": 0.949468085106383}, "shuffled rhyming lines": {"n": 336, "mean": 0.7529761904761905}, "shuffled rhyming words": {"n": 366, "mean": 0.912568306010929}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 376

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / shuffled rhyming lines / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.shuffled rhyming lines.mean; whole detail record: {"field_accuracy": 0.9228, "case_exact_accuracy": 0.6707768187422934, "custom_metrics": {}, "subgroups": {"beam search": {"n": 401, "mean": 0.830423940149626}, "corrupted rhyme": {"n": 376, "mean": 0.9840425531914894}, "delayed beam search": {"n": 137, "mean": 0.9927007299270073}, "delete any 1 nonrhyming word": {"n": 376, "mean": 0.8962765957446809}, "delete any 1 rhyming EOL word": {"n": 376, "mean": 0.9893617021276596}, "delete any 1 word": {"n": 376, "mean": 0.8962765957446809}, "delete any 2 nonrhyming words": {"n": 376, "mean": 0.9148936170212766}, "delete any 2 rhyming EOL words": {"n": 376, "mean": 0.9920212765957447}, "delete any 2 words": {"n": 376, "mean": 0.9468085106382979}, "delete any 3 nonrhyming words": {"n": 376, "mean": 0.9095744680851063}, "delete any 3 rhyming words": {"n": 376, "mean": 0.9840425531914894}, "delete any 3 words": {"n": 376, "mean": 0.949468085106383}, "shuffled rhyming lines": {"n": 336, "mean": 0.7529761904761905}, "shuffled rhyming words": {"n": 366, "mean": 0.912568306010929}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 336

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / shuffled rhyming words / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.shuffled rhyming words.mean; whole detail record: {"field_accuracy": 0.9228, "case_exact_accuracy": 0.6707768187422934, "custom_metrics": {}, "subgroups": {"beam search": {"n": 401, "mean": 0.830423940149626}, "corrupted rhyme": {"n": 376, "mean": 0.9840425531914894}, "delayed beam search": {"n": 137, "mean": 0.9927007299270073}, "delete any 1 nonrhyming word": {"n": 376, "mean": 0.8962765957446809}, "delete any 1 rhyming EOL word": {"n": 376, "mean": 0.9893617021276596}, "delete any 1 word": {"n": 376, "mean": 0.8962765957446809}, "delete any 2 nonrhyming words": {"n": 376, "mean": 0.9148936170212766}, "delete any 2 rhyming EOL words": {"n": 376, "mean": 0.9920212765957447}, "delete any 2 words": {"n": 376, "mean": 0.9468085106382979}, "delete any 3 nonrhyming words": {"n": 376, "mean": 0.9095744680851063}, "delete any 3 rhyming words": {"n": 376, "mean": 0.9840425531914894}, "delete any 3 words": {"n": 376, "mean": 0.949468085106383}, "shuffled rhyming lines": {"n": 336, "mean": 0.7529761904761905}, "shuffled rhyming words": {"n": 366, "mean": 0.912568306010929}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 366

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary cluster_macro_accuracy on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.cluster_macro_accuracy; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2000

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / A#7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.A#7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / A#M / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.A#M.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 88

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / A#m / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.A#m.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 43

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / A#m7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.A#m7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 7

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / A#maj7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.A#maj7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 3

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / A7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.A7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 4

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / AM / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.AM.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 73

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / Am / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.Am.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 69

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / Am7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.Am7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 8

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / Amaj7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.Amaj7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 3

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / B7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.B7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / BM / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.BM.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 88

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / Bdim / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.Bdim.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / Bhalf-dim7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.Bhalf-dim7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / Bm / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.Bm.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 65

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / Bm7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.Bm7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 6

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / Bmaj7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.Bmaj7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 3

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / C#7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.C#7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / C#M / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.C#M.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 78

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / C#aug / Faug / Aaug / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.C#aug / Faug / Aaug.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / C#dim / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.C#dim.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / C#m / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.C#m.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 51

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / C#m7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.C#m7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 8

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / C#maj7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.C#maj7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 6

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / C#sus2 / G#sus4 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.C#sus2 / G#sus4.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 4

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / C#sus4 / F#sus2 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.C#sus4 / F#sus2.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 6

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / C7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.C7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / CM / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.CM.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 99

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / Cdim / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.Cdim.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / Cm / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.Cm.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 67

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / Cm7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.Cm7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 4

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / Cmaj7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.Cmaj7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / Csus2 / Gsus4 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.Csus2 / Gsus4.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 9

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / Csus4 / Fsus2 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.Csus4 / Fsus2.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 6

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / D#7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.D#7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 7

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / D#M / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.D#M.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 86

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / D#aug / Gaug / Baug / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.D#aug / Gaug / Baug.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / D#dim / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.D#dim.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / D#m / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.D#m.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 43

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / D#m7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.D#m7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 3

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / D#maj7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.D#maj7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 6

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / D#sus2 / A#sus4 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.D#sus2 / A#sus4.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 7

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / D#sus4 / G#sus2 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.D#sus4 / G#sus2.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 4

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / D7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.D7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / DM / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.DM.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 77

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / Dm / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.Dm.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 64

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / Dm7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.Dm7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 8

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / DmMaj7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.DmMaj7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 1

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / Dmaj7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.Dmaj7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 3

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / Dsus2 / Asus4 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.Dsus2 / Asus4.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 12

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / Dsus4 / Gsus2 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.Dsus4 / Gsus2.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 5

Snapshot: 2026-10-12T08:15:19+00:00 · Revision: c5b04fe204d499ad0f65fa84b62b112922c18886 · Apache-2.0 published summary; source datasets retain their own terms

Publisher auxiliary subgroups / E7 / mean on this exact vendor public run. Not interchangeable with the primary native score.

Vendor-run Decision Index 0.3 public suite, kit apolinario/decision-index@62d2f51; 140178 requests, 43 benchmarks (37 in index), 2026-10-10, 2x NVIDIA A100 80GB PCIe, two length-sorted batching lanes; three overlong requests split across both GPUs. Base Gemma-4-31B-it@842da3794eaa0b77d5f08bae87a17459d91ff475 plus unmerged Sieve-31B LoRA@851574c0477d3286152965cf5161ce1957801db0 (same LFS adapter hash as run). BF16; logits/probabilities recorded at T=1 and re-read to training T=8.5495; decider-ai 1.6.0 chat answer slots; no truncation, no unsupported/error. Batch wall times set to zero and excluded from speed evidence. This larger public submission population is separate from the community 0.3.2 display and private Full score.

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers. Exact auxiliary path detail.subgroups.E7.mean; whole detail record: {"field_accuracy": 0.2795, "case_exact_accuracy": 0.2795, "cluster_macro_accuracy": 0.26706950456950457, "custom_metrics": {}, "subgroups": {"A#7": {"n": 1, "mean": 0}, "A#M": {"n": 88, "mean": 0.06818181818181818}, "A#m": {"n": 43, "mean": 0}, "A#m7": {"n": 7, "mean": 0.5714285714285714}, "A#maj7": {"n": 3, "mean": 0}, "A7": {"n": 4, "mean": 0}, "AM": {"n": 73, "mean": 0.136986301369863}, "Am": {"n": 69, "mean": 0.057971014492753624}, "Am7": {"n": 8, "mean": 0}, "Amaj7": {"n": 3, "mean": 0}, "B7": {"n": 2, "mean": 0}, "BM": {"n": 88, "mean": 0.011363636363636364}, "Bdim": {"n": 1, "mean": 0}, "Bhalf-dim7": {"n": 1, "mean": 0}, "Bm": {"n": 65, "mean": 0.18461538461538463}, "Bm7": {"n": 6, "mean": 0}, "Bmaj7": {"n": 3, "mean": 0}, "C#7": {"n": 1, "mean": 1}, "C#M": {"n": 78, "mean": 0.5897435897435898}, "C#aug / Faug / Aaug": {"n": 1, "mean": 1}, "C#dim": {"n": 1, "mean": 0}, "C#m": {"n": 51, "mean": 0.4117647058823529}, "C#m7": {"n": 8, "mean": 0.625}, "C#maj7": {"n": 6, "mean": 0}, "C#sus2 / G#sus4": {"n": 4, "mean": 0}, "C#sus4 / F#sus2": {"n": 6, "mean": 0.8333333333333334}, "C7": {"n": 2, "mean": 0.5}, "CM": {"n": 99, "mean": 0.8181818181818182}, "Cdim": {"n": 1, "mean": 0}, "Cm": {"n": 67, "mean": 0.8208955223880597}, "Cm7": {"n": 4, "mean": 0.25}, "Cmaj7": {"n": 1, "mean": 0}, "Csus2 / Gsus4": {"n": 9, "mean": 0.3333333333333333}, "Csus4 / Fsus2": {"n": 6, "mean": 0.16666666666666666}, "D#7": {"n": 7, "mean": 0}, "D#M": {"n": 86, "mean": 0.11627906976744186}, "D#aug / Gaug / Baug": {"n": 1, "mean": 0}, "D#dim": {"n": 1, "mean": 0}, "D#m": {"n": 43, "mean": 0.27906976744186046}, "D#m7": {"n": 3, "mean": 0}, "D#maj7": {"n": 6, "mean": 0}, "D#sus2 / A#sus4": {"n": 7, "mean": 0}, "D#sus4 / G#sus2": {"n": 4, "mean": 0}, "D7": {"n": 5, "mean": 0}, "DM": {"n": 77, "mean": 0.7402597402597403}, "Dm": {"n": 64, "mean": 0.53125}, "Dm7": {"n": 8, "mean": 0}, "DmMaj7": {"n": 1, "mean": 0}, "Dmaj7": {"n": 3, "mean": 0}, "Dsus2 / Asus4": {"n": 12, "mean": 0.16666666666666666}, "Dsus4 / Gsus2": {"n": 5, "mean": 0.2}, "E7": {"n": 2, "mean": 0}, "EM": {"n": 80, "mean": 0.25}, "Edim": {"n": 1, "mean": 0}, "Em": {"n": 74, "mean": 0.5}, "Em7": {"n": 6, "mean": 0.16666666666666666}, "Emaj7": {"n": 1, "mean": 0}, "Esus2 / Bsus4": {"n": 6, "mean": 0.16666666666666666}, "Esus4 / Asus2": {"n": 8, "mean": 0}, "F#7": {"n": 2, "mean": 0}, "F#M": {"n": 62, "mean": 0.43548387096774194}, "F#dim": {"n": 1, "mean": 0}, "F#half-dim7": {"n": 1, "mean": 0}, "F#m": {"n": 58, "mean": 0.15517241379310345}, "F#m7": {"n": 9, "mean": 0.2222222222222222}, "F#mMaj7": {"n": 1, "mean": 0}, "F#maj7": {"n": 4, "mean": 0}, "F#sus4 / Bsus2": {"n": 5, "mean": 0}, "F7": {"n": 2, "mean": 0}, "FM": {"n": 93, "mean": 0.043010752688172046}, "Fm": {"n": 55, "mean": 0.16363636363636364}, "Fm7": {"n": 7, "mean": 0}, "Fmaj7": {"n": 4, "mean": 0}, "Fsus4 / A#sus2": {"n": 9, "mean": 0}, "G#7": {"n": 6, "mean": 0}, "G#M": {"n": 88, "mean": 0}, "G#m": {"n": 50, "mean": 0}, "G#m7": {"n": 4, "mean": 0.25}, "G#maj7": {"n": 6, "mean": 0}, "G7": {"n": 2, "mean": 0}, "GM": {"n": 102, "mean": 0.3137254901960784}, "Gdim": {"n": 1, "mean": 0}, "Gm": {"n": 66, "mean": 0.4090909090909091}, "Gm7": {"n": 7, "mean": 0}, "Gmaj7": {"n": 2, "mean": 0}, "NoChord": {"n": 29, "mean": 0.5172413793103449}, "Other": {"n": 3, "mean": 0}}}

Vendor run over 140178 requests/43 benchmarks, not the board Full composite. The Decision Index 0.3 public index averages 37 benchmarks in five areas with fixed weights (knowledge 25.8%, language 25.8%, retrieval 20.0%, tools 18.3%, arts 10%). Inside an area, gold ★ benchmarks weigh 1.2 and the rest 1.0; the index is 100 x the weighted mean of the five areas. Each benchmark is chance-corrected first, (score - chance) / (1 - chance) clipped to 0-1, so 0 means random guessing and 100 means perfect. Every score is coverage-adjusted, so an unanswered or unsupported request counts as wrong. On the board this public index is 20% of the Full score; the rest comes from private tests run by the maintainers.

Checked / reported: 2026-10-20 · Evaluated: 2026-10-10 · Samples: 2